WO2022099526A1 - 变道预测回归模型训练方法、变道预测方法和装置 - Google Patents

变道预测回归模型训练方法、变道预测方法和装置 Download PDF

Info

Publication number
WO2022099526A1
WO2022099526A1 PCT/CN2020/128239 CN2020128239W WO2022099526A1 WO 2022099526 A1 WO2022099526 A1 WO 2022099526A1 CN 2020128239 W CN2020128239 W CN 2020128239W WO 2022099526 A1 WO2022099526 A1 WO 2022099526A1
Authority
WO
WIPO (PCT)
Prior art keywords
lane change
sample
target
probability
motion feature
Prior art date
Application number
PCT/CN2020/128239
Other languages
English (en)
French (fr)
Inventor
许家妙
何明
叶茂胜
曹通易
Original Assignee
深圳元戎启行科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳元戎启行科技有限公司 filed Critical 深圳元戎启行科技有限公司
Priority to CN202080092993.3A priority Critical patent/CN114945961B/zh
Priority to PCT/CN2020/128239 priority patent/WO2022099526A1/zh
Publication of WO2022099526A1 publication Critical patent/WO2022099526A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/16—Anti-collision systems
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Definitions

  • the present application relates to a lane change prediction regression model training method, lane change prediction method, device, computer equipment and storage medium.
  • Self-driving cars are intelligent cars that realize unmanned driving through computer systems.
  • the computer system automatically and safely controls the driving of the car without the active operation of the human being.
  • it is necessary to detect the motion of moving objects on the way or near the driving path to avoid the moving objects, thereby ensuring the safety of the autonomous vehicle.
  • the inventor realizes that the current method for recognizing the motion state of a moving object is inaccurate, resulting in low obstacle avoidance capability of the autonomous vehicle, and thus low safety of the autonomous vehicle.
  • a lane change prediction regression model training method a lane change prediction method, an apparatus, a computer device, and a storage medium.
  • a lane change prediction regression model training method comprising:
  • the first sample motion feature is obtained according to the motion data of the sample motion object collected at the sample collection time;
  • the standard lane change probability is based on the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time It is determined that the standard lane change probability is negatively correlated with the lane change time interval;
  • the model parameters in the lane change prediction regression model are adjusted by using the model loss value to obtain the trained lane change prediction regression model, so as to perform lane change prediction according to the trained lane change prediction regression model.
  • a lane change prediction regression model training device comprising:
  • a first sample motion feature acquisition module configured to acquire a first sample motion feature, the first sample motion feature is obtained according to the motion data of the sample motion object collected at the sample collection time;
  • the standard lane change probability acquisition module is used to obtain the standard lane change probability corresponding to the motion feature of the first sample, and the standard lane change probability is based on the object lane change time corresponding to the first sample motion feature and the sample A lane change time interval between acquisition times is determined, and the standard lane change probability is negatively correlated with the lane change time interval;
  • a predicted lane change probability obtaining module used for inputting the first sample motion feature into the lane change prediction regression model to be trained, to obtain the predicted lane change probability corresponding to the first sample motion feature;
  • a model loss value obtaining module for obtaining a model loss value according to the probability difference between the predicted lane change probability and the standard lane change probability
  • the trained lane change prediction regression model obtaining module is used to adjust the model parameters in the lane change prediction regression model by using the model loss value to obtain the trained lane change prediction regression model, so as to obtain the trained lane change prediction regression model according to the trained lane change prediction regression model.
  • the lane change prediction regression model is used for lane change prediction.
  • a computer device comprising a memory and one or more processors, the memory having computer-readable instructions stored therein, the computer-readable instructions, when executed by the processor, cause the one or more processors to execute The following steps:
  • the first sample motion feature is obtained according to the motion data of the sample motion object collected at the sample collection time;
  • the standard lane change probability is based on the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time It is determined that the standard lane change probability is negatively correlated with the lane change time interval;
  • the model parameters in the lane change prediction regression model are adjusted by using the model loss value to obtain the trained lane change prediction regression model, so as to perform lane change prediction according to the trained lane change prediction regression model.
  • One or more computer storage media storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the following steps:
  • the first sample motion feature is obtained according to the motion data of the sample motion object collected at the sample collection time;
  • the standard lane change probability is based on the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time It is determined that the standard lane change probability is negatively correlated with the lane change time interval;
  • the model parameters in the lane change prediction regression model are adjusted by using the model loss value to obtain the trained lane change prediction regression model, so as to perform lane change prediction according to the trained lane change prediction regression model.
  • a lane change prediction method comprising:
  • the target motion feature is input into the trained lane change prediction regression model to obtain the target lane change probability corresponding to the target moving object; the trained lane change prediction regression model is based on the first sample motion feature and the The standard lane change probability corresponding to the first sample motion feature is obtained by training, the first sample motion feature is obtained according to the motion data of the sample moving object collected at the sample collection time, and the standard lane change probability is obtained according to the determining the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time, and the standard lane change probability and the lane change time interval are negatively correlated; and
  • a target lane change result corresponding to the target moving object is determined according to the target lane change probability.
  • a lane change prediction device comprising:
  • the target motion feature acquisition module is used to acquire the target motion feature corresponding to the target motion object
  • the target lane change probability obtaining module is used to input the target motion feature into the trained lane change prediction regression model to obtain the target lane change probability corresponding to the target moving object;
  • the trained lane change prediction regression model Obtained by training according to the first sample motion feature and the standard lane change probability corresponding to the first sample motion feature, and the first sample motion feature is obtained according to the motion data of the sample moving object collected at the sample collection time , the standard lane change probability is determined according to the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time, the standard lane change probability and the lane change time interval negatively correlated;
  • a target lane change result determination module configured to determine the target lane change result corresponding to the target moving object according to the target lane change probability.
  • a computer device comprising a memory and one or more processors, the memory having computer-readable instructions stored therein, the computer-readable instructions, when executed by the processor, cause the one or more processors to execute The following steps:
  • the target motion feature is input into the trained lane change prediction regression model to obtain the target lane change probability corresponding to the target moving object; the trained lane change prediction regression model is based on the first sample motion feature and the The standard lane change probability corresponding to the first sample motion feature is obtained by training, the first sample motion feature is obtained according to the motion data of the sample moving object collected at the sample collection time, and the standard lane change probability is obtained according to the determining the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time, and the standard lane change probability and the lane change time interval are negatively correlated; and
  • a target lane change result corresponding to the target moving object is determined according to the target lane change probability.
  • One or more computer storage media storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the following steps:
  • the target motion feature is input into the trained lane change prediction regression model to obtain the target lane change probability corresponding to the target moving object; the trained lane change prediction regression model is based on the first sample motion feature and the The standard lane change probability corresponding to the first sample motion feature is obtained by training, the first sample motion feature is obtained according to the motion data of the sample moving object collected at the sample collection time, and the standard lane change probability is obtained according to the determining the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time, and the standard lane change probability and the lane change time interval are negatively correlated; and
  • a target lane change result corresponding to the target moving object is determined according to the target lane change probability.
  • FIG. 1 is an application scenario diagram of a lane change prediction regression model training method according to one or more embodiments
  • FIG. 2 is a schematic flowchart of a method for training a lane change prediction regression model according to one or more embodiments
  • FIG. 3 is a schematic flowchart of steps of obtaining a target time interval threshold and a target lane change probability generation factor according to one or more embodiments;
  • FIG. 4 is a schematic flowchart of a lane change prediction method according to one or more embodiments.
  • FIG. 5 is a block diagram of an apparatus for training a lane change prediction regression model according to one or more embodiments
  • FIG. 6 is a block diagram of an apparatus for lane change prediction according to one or more embodiments.
  • FIG. 7 is a block diagram of a computer device in accordance with one or more embodiments.
  • FIG. 8 is a block diagram of a computer device in accordance with one or more embodiments.
  • the lane change prediction regression model training method provided by the present application can be applied to the application environment shown in FIG. 1 .
  • the terminal 102 communicates with the server 104 through a network.
  • the server 104 obtains the first sample motion feature, which is obtained according to the motion data of the sample moving object collected at the sample collection time, and obtains the standard lane change probability corresponding to the first sample motion feature, and the standard change
  • the lane change probability is determined according to the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time, and the standard lane change probability is negatively correlated with the lane change time interval.
  • the predicted lane change probability corresponding to the motion feature of the first sample is obtained, and the model loss value is obtained according to the probability difference between the predicted lane change probability and the standard lane change probability, and the model loss value is used.
  • the model parameters in the lane change prediction regression model are adjusted to obtain a trained lane change prediction regression model, so as to perform lane change prediction according to the trained lane change prediction regression model.
  • the server 104 can use the trained lane change prediction regression model to perform lane change prediction, and control the movement according to the result of the lane change prediction.
  • the server 104 can also transmit the trained lane change prediction regression model to the terminal 102, and the terminal 102 can use the trained lane change prediction regression model.
  • the lane change prediction regression model is used to predict the lane change, and the server or the terminal controls the terminal to move according to the result of the lane change prediction.
  • the terminal 102 can be, but is not limited to, an autonomous vehicle and a mobile robot, and the server 104 can be implemented by an independent server or a server cluster composed of multiple servers.
  • a method for training a lane change prediction regression model is provided, and the method is applied to the server 104 in FIG. 1 as an example for description, including the following steps:
  • S202 Obtain a first sample motion feature, where the first sample motion feature is obtained according to motion data of the sample motion object collected at the sample collection time.
  • a moving object refers to an object in a moving state.
  • the object may be a living object, for example, may include at least one of a human being or an animal, or an inanimate object, such as at least one of a vehicle, an aircraft, or a bicycle.
  • the sample moving object can be any moving object or a specified moving object.
  • the motion data may be data related to the motion of the sample moving object collected at the time of sample collection, may include point cloud data collected by a point cloud collection device during the movement of the sample moving object, or may include data collected by an image collection device. at least one of the received image data.
  • a point cloud refers to a collection of three-dimensional data points in a three-dimensional coordinate system. For example, it can be a collection of three-dimensional data points corresponding to the surface of an object in a three-dimensional coordinate system.
  • a point cloud can represent the outer surface shape of an object.
  • Three-dimensional data points refer to points in three-dimensional space.
  • the three-dimensional data points may also include at least one of RGB color, grayscale value, or time.
  • the point cloud can be obtained by scanning with lidar.
  • the point cloud collection device can be any device that can collect point cloud data, and can be, but not limited to, a lidar, for example, a lidar set on the top of an autonomous vehicle.
  • the image acquisition device may be any device that can acquire image data, and may be, but not limited to, a camera.
  • Lidar is an active sensor that emits a laser beam, hits the laser beam on the surface of the object, the laser beam is bounced, and the bounced laser signal is collected to obtain the point cloud of the object.
  • the motion data may be pre-stored in the server.
  • the motion data can be collected by the point cloud collection equipment or image collection equipment installed in the sample moving object, or it can be collected by the point cloud collection equipment or image collection equipment in the surrounding environment of the sample moving object.
  • the distance between objects is less than the distance threshold value collected by the device installed in the object.
  • Motion features refer to motion-related features that are calculated from motion data.
  • the motion feature may be calculated based on motion feature related data extracted from motion data.
  • the motion feature may be motion feature related data, or may be calculated based on motion feature related data corresponding to motion data at different times.
  • the server can select at least two point cloud frames from the point cloud data, extract motion feature-related data from each of the selected point cloud frames, perform data fitting on the motion feature-related data, obtain a fitting result, and fit the fitting
  • the result is used as a motion feature, for example, a statistical value of data related to the motion feature can be calculated to obtain the motion feature.
  • the statistical value may include at least one of a mean or a variance.
  • the motion feature-related data may be extracted from motion data through a neural network model.
  • the server can input the motion data into the trained motion feature-related data recognition model, and the motion feature-related data recognition model can process the motion data, such as convolution processing, to obtain motion feature-related data.
  • the motion feature-related data may include at least one of the position, speed, acceleration, distance, motion direction, or relative distance of the motion object, for example, may include at least one of the motion object's position, speed, acceleration, or motion direction in the world coordinate system.
  • the world coordinate system can be a three-dimensional coordinate system, and the position in the world coordinate system can be represented by (x, y, z), and x, y, and z are coordinate axes that are perpendicular to each other and intersect.
  • the speed may include at least one of the horizontal speed or the vertical speed of the moving object relative to the road centerline, where the horizontal speed refers to the speed in the direction parallel to the road centerline, and the vertical speed refers to the direction perpendicular to the road centerline. speed in the direction.
  • the direction of movement can be, for example, the direction of the vehicle.
  • the relative distance can be the distance between the sample moving object and the reference object, and the reference object can be an object in a stationary state, for example, it can be at least one of a road boundary or an object set on the road, and the object set on the road can be, for example, trees.
  • the reference object can also be a virtual thing, such as a route, which refers to the route of the aircraft, or a position in a three-dimensional coordinate system, such as the origin of the three-dimensional coordinate system.
  • Road boundaries can also be referred to as lane boundaries. That is, the relative distance may be the distance between the moving object and the road boundary, and the distance between the moving object and the road boundary may be referred to as the road boundary distance.
  • the road boundary distance may include at least one of a road left boundary distance or a road right boundary distance. The distance from the left border of the road refers to the distance between the sample moving object and the left border of the road, and the distance from the right border of the road refers to the distance between the sample moving object and the right border of the road.
  • the motion feature can also be determined according to the sorting result of the motion feature related data, and the sorting result of the motion feature related data can be used as the motion feature.
  • the relative distances at different times can be sorted in chronological order to obtain the distance sorting result, as Motion features, or the motion features are obtained by calculating the sorting results of the data related to the motion features.
  • the data in the distance sorting results can be normalized to obtain the motion features, or the data in the distance sorting results can be statistically calculated or calculated. Fitting to obtain motion features, or splicing the ranking results of motion feature-related data to obtain stitching results.
  • motion features for example, distance ranking results and speed ranking results can be spliced to obtain motion features.
  • the speed ranking results refer to different The sorting result obtained by sorting the speed of the moment in chronological order, for example, it can be the sorting result obtained by sorting the horizontal speed at different times in chronological order, for example, the distance sorting result is (s1, s2, s3), and the speed sorting result is (v1, v2, v3), then the result of splicing the distance sorting result and the speed sorting result can be expressed as (s1, s2, s3, v1, v2, v3).
  • the number of relative distances in the distance sorting result and the number of speeds in the speed sorting result may be consistent or inconsistent.
  • the point cloud frame corresponding to the relative distance in the distance sorting result may be consistent with the point cloud frame corresponding to the speed in the speed sorting result, or may be inconsistent.
  • the sample motion feature refers to the motion feature of the sample moving object.
  • the first sample motion feature can be any sample motion feature or a specified sample motion feature, which is used to train the lane change prediction regression model and obtain the trained model.
  • a lane change prediction regression model is a regression model for predicting the probability of a lane change occurring.
  • the input of the lane change prediction regression model can be motion features, and the output can be the probability of lane change.
  • the lane change prediction regression model can be an existing regression model or a custom regression model, such as a four-layer neural network. Regression models can also be referred to as regressors and recurrent neural networks.
  • the sample collection time refers to the time when the motion data is collected, which can be a time point or a time period, such as (t1, t2), where t1 represents the start time of the sample collection time, and t2 represents the start time of the sample collection time.
  • the termination time, the duration of the time period may be, for example, 5 minutes.
  • the server may acquire the motion point cloud of the sample moving object during the motion from the point cloud acquisition device, acquire the first number of point cloud frames from the point cloud frames included in acquiring the motion point cloud, and determine the first number
  • the motion feature-related data of the sample moving objects in each point cloud frame of the point cloud frame sort the motion feature-related data corresponding to each point cloud frame according to the time sequence of the point cloud frames, and obtain the sorting result, according to the sorting result
  • Obtain the sample motion characteristics such as calculating the statistical value of the data in the sorting result, obtain the sample motion characteristics, or perform data fitting on the data in the sorting result, and obtain the sample motion characteristics according to the data fitting results.
  • the result is used as the sample motion feature, or the result of data fitting is normalized to obtain the sample motion feature.
  • the point cloud frames in the first number of point cloud frames may be continuous or discontinuous point cloud frames.
  • the first number may be preset or set as required, for example, it may be 10.
  • S204 Obtain a standard lane change probability corresponding to the motion feature of the first sample, the standard lane change probability is determined according to the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample collection time, and the standard lane change probability There is a negative correlation with the lane change interval.
  • the object lane change time refers to the time when the sample moving object changes lanes.
  • the object lane change time corresponding to the first sample motion feature refers to the time when the sample moving object changes lanes for the first time after the sample acquisition time. For example, after the sample collection time (t1, t2), the sample moving object changes lanes at time t2+a1, then t2+a1 is the time for the object to change lanes.
  • the lane change time interval refers to the time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time.
  • the standard lane change probability corresponding to the motion feature of the first sample refers to the real lane change probability of the motion feature of the first sample.
  • the standard lane change probability can be negatively correlated with the lane change time interval.
  • the time interval condition may include that the lane change time interval is less than or equal to the target time interval threshold.
  • the target time interval threshold may be preset, or may be determined by training a lane change prediction regression model, for example, may be determined by cross-validation.
  • the standard lane change probability may be calculated according to the square value of the lane change time interval, for example, it may be calculated according to the ratio of the square value of the lane change time interval to a specific value
  • the obtained specific value may be preset or determined by training the lane change prediction regression model.
  • the server may obtain a fixed lane change probability as a standard lane change probability.
  • the fixed lane change probability may be set as required, or may be preset, for example, it may be 0.
  • S206 Input the motion feature of the first sample into the lane change prediction regression model to be trained to obtain a predicted lane change probability corresponding to the motion feature of the first sample.
  • the lane change prediction regression model to be trained refers to a lane change prediction regression model that needs to be trained.
  • the server may use the motion feature of the first sample as the input of the lane change prediction regression model to be trained, and the lane change prediction regression model to be trained may calculate the motion feature of the first sample, such as convolution calculation, to obtain the first sample
  • the predicted lane change probability is the lane change probability output by the lane change prediction regression model.
  • the probability difference refers to the difference between the predicted lane change probability and the standard lane change probability.
  • the model loss value is positively correlated with the probability difference.
  • the model loss value may be, for example, any one of the square of the probability difference or a multiple of the square of the probability difference, and the multiple is, for example, one-half.
  • the model loss value may be determined according to the magnitude relationship between the probability difference and the difference threshold. For example, when the probability difference is smaller than the difference threshold, the model loss value may be determined according to the square of the probability difference, and the model loss value is positively correlated with the square of the probability difference. When the probability difference is greater than or equal to the difference threshold, the model loss value can be calculated according to the probability difference and the difference threshold.
  • the difference threshold can be preset or set as required.
  • the model parameter refers to the variable parameter inside the lane change prediction regression model, and for the neural network model, it may also be referred to as a neural network weight (weight).
  • the trained lane change prediction regression model may be obtained through one or more trainings.
  • the server can adjust the model parameters in the lane change prediction regression model in the direction of decreasing the loss value, and can obtain the trained lane change prediction regression model after multiple iterations of training.
  • the server may perform backpropagation according to the model loss value, and in the process of backpropagation, update the model parameters of the lane change prediction regression model along the gradient descent direction to obtain the trained lane change prediction regression model.
  • reverse means that the direction of parameter update and lane change prediction is opposite.
  • the parameter update is back-propagated, the descending gradient can be obtained according to the model loss value, and the last layer of the regression model can be predicted from the lane change.
  • the gradient update of the model parameters is started according to the descending gradient, until the first layer of the lane change prediction regression model is reached.
  • the gradient descent method can be either stochastic gradient descent or batch gradient descent.
  • the training of the model can be iterated multiple times, that is, the trained lane change prediction regression model can be obtained by iterative training, and the training is stopped when the model convergence condition is satisfied. Set the loss value, or it can be that the change of the model parameters is smaller than the preset parameter change value.
  • the server may transmit the trained lane change prediction regression model to the terminal, and the terminal uses the trained lane change prediction regression model to predict the lane change of the moving objects in the surrounding environment, thereby controlling the movement of the terminal, realizing Avoidance.
  • a first sample motion feature is obtained, and the first sample motion feature is obtained according to the motion data of the sample moving object collected at the sample collection time, and the obtained first sample motion feature corresponds to.
  • the standard lane-change probability of The probability difference between the two is obtained, the model loss value is obtained, the model parameters in the lane change prediction regression model are adjusted by the model loss value, and the trained lane change prediction regression model is obtained, and the lane change prediction is performed according to the trained lane change prediction regression model.
  • the standard lane change probability is determined according to the lane change time interval between the object lane change time corresponding to the first sample motion feature and the sample acquisition time, the standard lane change probability is negatively correlated with the lane change time interval, thereby improving the trained
  • the accuracy of the regression model of lane change prediction improves the accuracy of lane change prediction.
  • obtaining the standard lane change probability corresponding to the motion feature of the first sample includes: when the lane change time interval is less than or equal to the target time interval threshold, obtaining the target lane change probability generation factor; and, according to the lane change time interval and the target lane change probability generation factor to calculate the standard lane change probability corresponding to the motion feature of the first sample.
  • the time interval threshold may be a period of time, such as 10 seconds.
  • the target time interval threshold may be selected from candidate time interval thresholds.
  • the candidate time interval threshold may be a time interval threshold pre-stored in the server, or may be a time interval threshold acquired or generated by the server.
  • the lane change probability generation factor is used to generate the lane change probability.
  • the lane change probability refers to the probability of the moving object changing lanes.
  • the target lane change probability generation factor may be selected from the candidate lane change probability generation factors.
  • the candidate lane change probability generation factor may be a lane change probability generation factor pre-stored in the server, or may be a lane change probability generation factor acquired or generated by the server.
  • the target lane change probability generation factor may be determined by training a lane change prediction regression model.
  • the server may generate the candidate lane change probability according to the candidate lane change probability generation factor and the lane change time interval, use the candidate lane change probability as the probability label of the motion feature of the second sample, and use the motion feature of the second sample and the corresponding probability label to compare the
  • the lane change prediction regression model is trained, and the target lane change probability generation factor is determined according to the accuracy of the trained lane change prediction regression model.
  • the second sample motion feature may be any sample motion feature, or may be a specified sample motion feature, for example, may be a sample motion feature different from the first sample motion feature. Probabilistic labels can also be called regression ground truth.
  • the server may use a lane change probability calculation formula to calculate the lane change time interval and the target lane change probability generation factor to obtain the standard lane change probability corresponding to the motion feature of the first sample.
  • the calculation formula of lane change probability when the value of the independent variable is greater than or equal to the value of the independent variable, the dependent variable decreases with the increase of the independent variable, and the value of the dependent variable is greater than 0 and less than 1.
  • the server can replace the independent variable in the lane change probability calculation formula with the lane change time interval, replace the constant in the lane change probability calculation formula with the target lane change probability generation factor, and calculate the standard lane change corresponding to the motion feature of the first sample. probability.
  • the calculation formula of the lane change probability may be, for example, a Gaussian function.
  • the server may use a Gaussian function to calculate the lane change time interval and the target lane change probability generation factor to obtain the standard lane change probability corresponding to the motion feature of the first sample.
  • the server can replace the independent variable of the Gaussian function with the lane change time interval, replace the variance in the Gaussian function with the target lane change probability generation factor, set the mean value of the Gaussian function to 0, and calculate the corresponding motion feature of the first sample.
  • Standard lane change probability is used to calculate the lane change time interval.
  • the target lane change probability generation factor is obtained, and the calculation is performed according to the lane change time interval and the target lane change probability generation factor, so as to obtain the corresponding motion characteristics of the first sample.
  • the standard lane change probability can be ensured that the lane change time interval corresponding to the standard lane change probability is not too large, the validity of the lane change time interval is ensured, and the accuracy of the standard lane change probability is improved.
  • calculating according to the lane change time interval and the target lane change probability generation factor to obtain the standard lane change probability corresponding to the motion feature of the first sample includes: performing a square operation on the lane change time interval to obtain a squared value of the interval; Calculate the ratio of the interval squared value to the target lane change probability generation factor to obtain the target ratio; and, take the first value as the base, and use the negative number of the target ratio as the index to perform exponential calculation to obtain the standard lane change corresponding to the motion characteristic of the first sample probability.
  • the squared value of the interval is the square of the lane-change time interval.
  • the squared value of the interval is d 2 .
  • the target ratio refers to the ratio of the interval squared value to the target lane change probability generation factor, for example, it may be d 2 /C, where C represents the target lane change probability generation factor.
  • the first value can be any positive number greater than 1, for example, it can be the base of the natural logarithm e ⁇ 2.71828.
  • the server may use the first value as a base, and perform index calculation according to the target ratio determination index to obtain the standard lane change probability.
  • the server may use the first value as the base, and use the negative number of the target ratio as the index to perform exponential calculation to obtain the standard lane change probability corresponding to the motion feature of the first sample.
  • the server may use the first value as the base, and use the ratio of the negative number of the target ratio to the second value as an index to perform exponential calculation to obtain the standard lane change probability.
  • the second value can be any positive number, for example, it can be 2.
  • the standard lane change probability can be expressed as formula (1), for example, where c 2 represents the target lane change probability generation factor. f(d) represents the standard lane change probability.
  • a square operation is performed on the lane change time interval to obtain the interval square value, the ratio of the interval square value to the target lane change probability generation factor is calculated, and the target ratio value is obtained, the first value is used as the base, and the negative number of the target ratio value is used as The index performs exponential calculation to obtain the standard lane change probability corresponding to the motion feature of the first sample, so that the value of the standard lane change probability can be guaranteed to be between 0 and 1, and the accuracy of the standard lane change probability is improved.
  • the steps of obtaining the target time interval threshold and the target lane change probability generation factor include:
  • S302 Determine at least two candidate parameter combinations, where the candidate parameter combinations include a candidate time interval threshold and a candidate lane change probability generation factor.
  • the server can arbitrarily select a candidate time interval threshold from a plurality of candidate time interval thresholds, as a candidate time interval threshold in the candidate parameter combination, and can arbitrarily select a candidate lane change from a plurality of candidate lane change probability generation factors Probability generation factor as the candidate interval threshold in the candidate parameter combination.
  • the number of candidate time interval thresholds is N
  • the number of candidate lane change probability generation factors is M
  • the number of candidate parameter combinations may be N ⁇ M.
  • the candidate time interval thresholds are r1 and r2
  • the candidate lane change probability generation factors are w1 and w2
  • the candidate parameter combinations may include (r1,w1), (r1,w2), (r2,w1) and (r2,w2) ).
  • S304 Obtain the motion feature of the second sample, and determine the lane change probability corresponding to the motion feature of the second sample according to the candidate parameter combination, as a probability label corresponding to the motion feature of the second sample.
  • the second sample motion feature may be any sample motion feature, or may be a specified sample motion feature, for example, may be a sample motion feature different from the first sample motion feature.
  • the probability label refers to the label of the motion feature of the second sample, and the probability label can also be called the regression true value.
  • the probability label is obtained in a similar way to the standard lane change probability.
  • the probability label is determined according to the lane change time interval corresponding to the motion feature of the second sample. When the lane change time interval is less than or equal to the candidate time interval threshold in the candidate parameter combination, according to the lane change time interval The interval and the candidate lane change probability generation factor in the candidate parameter combination are calculated to obtain the probability label corresponding to the motion feature of the second sample. When the lane change time interval is greater than the candidate time interval threshold in the candidate parameter combination, the probability label corresponding to the motion feature of the second sample is determined as a fixed probability.
  • the fixed probability can be set as required or preset, such as 0.
  • the server may obtain a candidate sample motion feature set, and select a second number of candidate sample motion features from the candidate sample motion feature set to obtain a second sample motion feature set.
  • a third sample motion feature set is composed of candidate sample motion features other than the second sample motion feature set in the candidate sample motion feature set.
  • the second number can be set as required, for example, determined according to the number of candidate sample motion features in the candidate sample motion feature set (referred to as the total number of features), and the second number is, for example, one-fifth of the total number of features.
  • the server may input the motion feature of the second sample into the lane change prediction regression model to obtain the predicted lane change probability corresponding to the motion feature of the second sample, and the server may obtain the probability difference between the predicted lane change probability and the probability label according to the probability difference between the predicted lane change probability and the probability label
  • the model loss value is used to adjust the model parameters in the lane change prediction regression model by using the model loss value to obtain the trained lane change prediction regression model.
  • each second sample motion feature in the second sample motion feature set may be input into the lane change prediction regression model to obtain a trained lane change prediction regression model.
  • the server may obtain multiple second sample motion feature sets from the candidate sample motion feature sets, and the second sample motion features in different second sample motion feature sets are different. Using the same candidate parameter combination, the probability labels of the second sample motion features in each second sample motion feature set are determined. Each second sample motion feature set is used to train the lane change prediction regression model, respectively, to obtain the trained lane change prediction regression model obtained from different second sample motion feature sets under the same candidate parameter combination.
  • S308 Determine the model accuracy of the trained lane change prediction regression model corresponding to the candidate parameter combination, select a target parameter combination that satisfies the accuracy condition from at least two candidate parameter combinations based on the model accuracy, and select the target parameter combination in the target parameter combination.
  • the candidate time interval threshold is used as the target time interval threshold, and the candidate lane change probability generation factor in the target parameter combination is used as the target lane change probability generation factor.
  • the candidate parameter combination may further include a candidate lane change probability threshold.
  • the server may input the third sample motion feature in the third sample motion feature set into the trained lane change prediction regression model, obtain the predicted lane change probability of the third sample motion feature output by the lane change prediction regression model, and convert the predicted lane change
  • the lane change probability is compared with the candidate lane change probability threshold.
  • the predicted lane change probability is greater than the candidate lane change probability threshold, the predicted lane change result of the motion feature of the third sample is determined as a lane change, and when the predicted lane change probability is less than or equal to the candidate lane change
  • the probability threshold is set, it is determined that the predicted lane change result of the motion feature of the third sample is an unchanged lane.
  • the predicted lane change result is the predicted lane change result.
  • the server can obtain the standard lane change result of the motion feature of the third sample, compare the predicted lane change result with the standard lane change result, and when the comparison is consistent, determine the prediction result of the trained lane change prediction regression model on the motion feature of the third sample is correct, otherwise the prediction result is wrong.
  • the standard lane change result is the real lane change result of the motion feature of the third sample.
  • the server can determine the model accuracy of the trained lane change prediction regression model according to the prediction results corresponding to the motion characteristics of each third sample. For example, the model accuracy can be obtained by dividing the number of correct prediction results by the number of all prediction results. .
  • the accuracy condition may include at least one of accuracy greater than an accuracy threshold or model accuracy being at a maximum.
  • the accuracy threshold can be set as required, or can be preset, for example, it can be 95%.
  • the target parameter combination is a candidate parameter combination that satisfies the accuracy condition among the obtained at least two candidate
  • the server may determine the respective accuracies of the trained lane change prediction regression models obtained from different second sample motion feature sets under the same candidate parameter combination, and calculate the average value of the respective accuracies as the candidate parameter combination.
  • the model accuracy of the corresponding trained lane change prediction regression model The server may determine the model accuracy corresponding to each candidate parameter combination in the at least two candidate parameter combinations, and select a candidate parameter combination that satisfies the accuracy condition from the model accuracy corresponding to each candidate parameter combination as the target parameter combination, for example The candidate parameter combination corresponding to the maximum model accuracy is selected as the target parameter combination.
  • the server may use the candidate lane change probability threshold in the target parameter combination as the target lane change probability threshold.
  • the server may combine the target parameter combination with the corresponding trained lane change prediction regression model as the lane change prediction regression model to be trained.
  • the model accuracy of the trained lane change prediction regression model corresponding to the candidate parameter combination is determined, and the target parameter combination that satisfies the accuracy condition is selected from at least two candidate parameter combinations based on the model accuracy, and the target parameter
  • the candidate time interval threshold in the combination is used as the target time interval threshold, and the candidate lane change probability generation factor in the target parameter combination is used as the target lane change probability generation factor, so as to obtain a target time interval that can make the lane change prediction regression model with high accuracy Threshold and target lane change probability generation factor.
  • the first sample motion feature includes a distance change feature
  • the step of determining the distance change feature includes: acquiring a target distance sequence corresponding to the sample moving object during the sample collection time, and the target distance sequence is sorted according to the motion time; The magnitude relationship of the distances in the distance sequence determines a distance change trend; and, according to the distance change trend, a distance change feature corresponding to the sample moving object is determined.
  • the distance sequence refers to a sequence composed of relative distances, and each relative distance in the distance sequence is sorted by movement time.
  • the earlier the movement time is, the earlier the relative distance is sorted.
  • it can be (s1, s2, s3, s4, s5, s6), where s1 to S6 respectively represent a relative distance.
  • the movement time refers to the time during the movement of the sample moving object, which can be a time point or a time period.
  • the sample moving object may correspond to multiple distance sequences, for example, may correspond to at least one of a left border distance sequence or a right border distance sequence.
  • the left border distance sequence refers to a sequence obtained by sorting the distances of the left border of the road in chronological order
  • the right border distance sequence refers to a sequence obtained by sorting the distances of the right border of the road in chronological order.
  • Distance sequences can also be represented by vectors.
  • the sample moving objects can correspond to multiple distance sequences.
  • the server can obtain multiple distance sequences corresponding to the sample moving objects according to the motion data collected at the sample collection time, and the target distance sequence can be from each distance sequence corresponding to the sample moving objects.
  • the selected distance sequence It can be selected by random selection, or can be selected by a preset selection method.
  • each distance sequence of the sample moving object can be respectively used as the target distance sequence.
  • the distance change trend refers to the law of the distance in the distance series changing with time, which can include either a gradually decreasing change trend or a gradually increasing change trend.
  • the distance change trend condition may be, for example, a gradually decreasing change trend.
  • the left border distance sequence is used as the target distance sequence.
  • the right The boundary distance sequence is used as the target distance sequence.
  • the distance change feature can be the feature of the distance change trend, and can be calculated according to the distance change trend.
  • the speed of the distance change can be determined according to the distance change trend to obtain the distance change feature.
  • the speed of the distance change may include at least one of the speed of the distance becoming larger or the speed of the distance becoming smaller.
  • the distance change feature can also be determined according to the average distance change degree, which refers to the size of the distance change per unit time, and can be calculated according to the target distance sequence and the time series corresponding to the target distance sequence.
  • the time sequence corresponding to the target distance sequence includes the time corresponding to each distance in the target distance sequence, and the time in the time sequence is sorted in order from front to back. The higher the time, the higher the sorting.
  • the time series may be (t1, t2, t3, t4, t5, t6), for example. Among them, t1 occurs before t6.
  • the distance difference can be obtained by the difference between the distance of the starting position and the distance of the ending position in the target distance sequence, and the time difference can be obtained according to the difference between the time of the starting position and the ending position in the time series.
  • the ratio to the time difference obtains the average distance change degree, and the average distance change degree can be, for example, (s6-s1) ⁇ (t6-t1).
  • the distance change feature can be represented by a vector, for example, it can be called a distance feature vector.
  • the distance change feature can also be a target distance sequence, that is, the target distance sequence can be used as the distance change feature.
  • the server may calculate the target distance sequence and the corresponding time sequence by using the change feature calculation method to obtain the distance transformation feature.
  • the variation characteristic calculation method may be, for example, formulas (2) and (3).
  • x represents the distance in the target distance sequence
  • y represents the time corresponding to the distance in the target distance sequence
  • a represents the distance change feature, which can also be understood as the change rate (slope) of the relative distance with time.
  • the target distance sequence corresponding to the sample moving object is obtained, the target distance sequence is sorted according to the movement time, the distance change trend is determined according to the size relationship of the distances in the target distance sequence, and the sample movement is determined according to the distance change trend.
  • the distance change feature corresponding to the object improves the accuracy of the motion feature of the first sample.
  • a method for predicting a lane change is provided, which is described by taking the method applied to the terminal 102 in FIG. 1 as an example, including the following steps:
  • the target moving object may be a moving object existing in the surrounding environment of the terminal.
  • the surrounding environment of the terminal may be, for example, an area surrounded by a circle with the terminal as the center and a fixed radius as the radius.
  • the fixed radius can be set as needed or preset.
  • the target motion feature refers to the motion feature of the current time corresponding to the target motion object.
  • the current time is the time period
  • the ending time of the current time is the current time
  • the starting time of the current time is the historical time whose time interval from the current time is the first time interval.
  • the first time interval may be preset or determined as required.
  • the trained lane change prediction regression model is based on the first sample motion feature and the first sample motion
  • the standard lane change probability corresponding to the feature is obtained by training, the first sample motion feature is obtained according to the motion data of the sample moving object collected at the sample collection time, and the standard lane change probability is obtained according to the object lane change corresponding to the first sample motion feature
  • the lane change time interval between the time and the sample collection time is determined, and the standard lane change probability is negatively correlated with the lane change time interval.
  • the trained lane change prediction regression model may be trained by using the above proposed lane change prediction regression model training method.
  • the first time interval may be consistent with the duration of the sample collection time.
  • S406 Determine a target lane change result corresponding to the target moving object according to the target lane change probability.
  • the target lane change result may include any one of lane change or unchanged lane, and may also include predicted lane change time information.
  • the predicted lane change time information is used to predict the time when the lane change occurs, for example, it may be within 5 seconds in the future. If the target lane change result is a lane change, it means that the target moving object is about to change lanes, for example, a lane change may occur in the next 5 seconds.
  • the lane change can be represented by 1, and the unchanged lane can be represented by 0.
  • the terminal may control the movement of the terminal according to the target lane change result, so as to avoid collision with the target moving object. For example, when the target lane change result of the moving object on the road in front of the terminal is a lane change, the terminal may decelerate and drive.
  • the target motion feature corresponding to the target moving object is obtained, and the target motion feature is input into the trained lane change prediction regression model to obtain the target lane change probability corresponding to the target moving object, which is determined according to the target lane change probability.
  • the target lane change result corresponding to the target moving object is obtained by training the trained lane change prediction regression model according to the first sample motion feature and the standard lane change probability corresponding to the first sample motion feature.
  • the standard lane change probability is determined according to the lane change time interval between the object lane change time corresponding to the first sample movement feature and the sample collection time, and the standard lane change probability is obtained from the motion data of the sample moving object collected at the sample collection time. There is a negative correlation with the lane change time interval, thereby improving the accuracy of the trained lane change prediction regression model and improving the lane change prediction accuracy.
  • determining the target lane change result corresponding to the target moving object according to the target lane change probability includes: comparing the target lane change probability with the target lane change probability threshold to obtain a comparison result; and determining the target movement according to the comparison result The target lane change result corresponding to the object.
  • the target lane change probability threshold may be preset or determined by training a lane change prediction regression model, for example, it may be 0.9.
  • the comparison result may be any one of the target lane change probability is greater than the target lane change probability threshold, and the target lane change probability is less than or equal to the target lane change probability threshold.
  • the terminal may determine that the target lane change result is a lane change; otherwise, the terminal may determine that the target lane change result is an unchanged lane.
  • the target lane change probability is compared with the target lane change probability threshold to obtain a comparison result, and the target lane change result corresponding to the target moving object is determined according to the comparison result, which improves the accuracy of the target lane change result.
  • steps in the flowcharts of FIGS. 2-4 are shown in sequence according to the arrows, these steps are not necessarily executed in the sequence shown by the arrows. Unless explicitly stated herein, the execution of these steps is not strictly limited to the order, and these steps may be performed in other orders. Moreover, at least a part of the steps in FIGS. 2-4 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. These sub-steps or stages are not necessarily completed at the same time. The order of execution of the steps is not necessarily sequential, but may be performed alternately or alternately with other steps or at least a part of sub-steps or stages of other steps.
  • a lane change prediction regression model training device including: a first sample motion feature acquisition module 502, a standard lane change probability acquisition module 504, and a predicted lane change probability acquisition module 506.
  • the model loss value obtaining module 508 and the trained lane change prediction regression model obtaining module 510 wherein:
  • the first sample motion feature acquisition module 502 is configured to acquire the first sample motion feature, and the first sample motion feature is obtained according to the motion data of the sample motion object collected at the sample collection time.
  • the standard lane change probability acquisition module 504 is used to obtain the standard lane change probability corresponding to the motion feature of the first sample, and the standard lane change probability is based on the lane change between the object lane change time corresponding to the first sample motion feature and the sample acquisition time The time interval is determined, and the standard lane change probability is negatively correlated with the lane change time interval.
  • the predicted lane change probability obtaining module 506 is configured to input the motion feature of the first sample into the lane change prediction regression model to be trained to obtain the predicted lane change probability corresponding to the motion feature of the first sample.
  • the model loss value obtaining module 508 is configured to obtain the model loss value according to the probability difference between the predicted lane change probability and the standard lane change probability.
  • the trained lane change prediction regression model obtaining module 510 is used to adjust the model parameters in the lane change prediction regression model by using the model loss value to obtain the trained lane change prediction regression model, so as to perform the lane change prediction regression model according to the trained lane change prediction regression model. Lane change prediction.
  • the standard lane change probability acquisition module 504 includes:
  • the target lane change probability generation factor obtaining unit is configured to obtain the target lane change probability generation factor when the lane change time interval is less than or equal to the target time interval threshold.
  • the standard lane change probability obtaining unit is used for calculating according to the lane change time interval and the target lane change probability generation factor to obtain the standard lane change probability corresponding to the motion feature of the first sample.
  • the standard lane change probability obtaining unit is further configured to perform a square operation on the lane change time interval to obtain the interval square value; calculate the ratio of the interval square value to the target lane change probability generation factor to obtain the target ratio value; and, The first value is used as the base, and the negative number of the target ratio is used as the index to perform exponential calculation to obtain the standard lane change probability corresponding to the motion feature of the first sample.
  • the lane change prediction regression model training device further includes a target time interval threshold obtaining module, and the target time interval threshold obtaining module includes:
  • the candidate parameter combination determination unit is configured to determine at least two candidate parameter combinations, where the candidate parameter combinations include a candidate time interval threshold and a candidate lane change probability generation factor.
  • the probability label obtaining unit is configured to obtain the motion feature of the second sample, and determine the lane change probability corresponding to the motion feature of the second sample according to the candidate parameter combination, as the probability label corresponding to the motion feature of the second sample.
  • the trained lane change prediction regression model obtaining unit is used to train the lane change prediction regression model based on the motion feature of the second sample and the probability label corresponding to the motion feature of the second sample, and obtain the trained lane change corresponding to the candidate parameter combination Predictive regression models.
  • the unit for obtaining the target time interval threshold is used to determine the model accuracy of the trained lane change prediction regression model corresponding to the candidate parameter combination, and based on the model accuracy, select the target parameter combination that satisfies the accuracy condition from at least two candidate parameter combinations , the candidate time interval threshold in the target parameter combination is used as the target time interval threshold, and the candidate lane change probability generation factor in the target parameter combination is used as the target lane change probability generation factor.
  • the first sample motion feature includes a distance change feature
  • the lane change prediction regression model training device further includes a distance change feature determination module
  • the distance change feature determination module includes:
  • the target distance sequence acquisition unit is used for acquiring the target distance sequence corresponding to the sample moving object in the sample collection time, and the target distance sequence is sorted according to the movement time.
  • the distance change trend determination unit is used for determining the distance change trend according to the magnitude relationship of the distances in the target distance sequence.
  • the distance change feature determination unit is used for determining the distance change feature corresponding to the sample moving object according to the distance change trend.
  • Each module in the above-mentioned lane change prediction regression model training device may be implemented in whole or in part by software, hardware and combinations thereof.
  • the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • a lane change prediction apparatus including: a target motion feature acquisition module 602, a target lane change probability acquisition module 604, and a target lane change result determination module 606, wherein:
  • the target motion feature acquisition module 602 is configured to acquire target motion features corresponding to the target motion object.
  • the target lane change probability obtaining module 604 is used to input the target motion feature into the trained lane change prediction regression model to obtain the target lane change probability corresponding to the target moving object; the trained lane change prediction regression model is based on the first sample
  • the motion feature and the standard lane change probability corresponding to the motion feature of the first sample are obtained by training, the first sample motion feature is obtained according to the motion data of the sample moving object collected at the sample collection time, and the standard lane change probability is obtained according to the first sample
  • the lane-change time interval between the object's lane-change time corresponding to this motion feature and the sample collection time is determined, and the standard lane-change probability has a negative correlation with the lane-change time interval.
  • the target lane change result determination module 606 is configured to determine the target lane change result corresponding to the target moving object according to the target lane change probability.
  • the target lane change result determination module 606 includes:
  • the comparison result obtaining unit is used for comparing the target lane change probability with the target lane change probability threshold to obtain the comparison result.
  • the target lane change result obtaining unit is used for determining the target lane change result corresponding to the target moving object according to the comparison result.
  • Each module in the above-mentioned lane change prediction apparatus may be implemented in whole or in part by software, hardware and combinations thereof.
  • the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • a computer device is provided, and the computer device may be a server, and its internal structure diagram may be as shown in FIG. 7 .
  • the computer device includes a processor, memory, a network interface, and a database connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium, an internal memory.
  • the non-volatile storage medium stores an operating system, computer readable instructions and a database.
  • the internal memory provides an environment for the execution of the operating system and computer-readable instructions in the non-volatile storage medium.
  • the database of the computer equipment is used to store data such as sample motion characteristics, standard lane change probability, predicted lane change probability, model loss value, and lane change prediction regression model.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer readable instructions when executed by a processor, implement a method for training a lane change prediction regression model.
  • a computer device is provided, and the computer device may be a terminal, and its internal structure diagram may be as shown in FIG. 8 .
  • the computer equipment includes a processor, memory, a communication interface, a display screen, and an input device connected by a system bus.
  • the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium, an internal memory.
  • the nonvolatile storage medium stores an operating system and a computer program.
  • the internal memory provides an environment for the execution of the operating system and computer programs in the non-volatile storage medium.
  • the communication interface of the computer device is used for wired or wireless communication with an external terminal, and the wireless communication can be realized by WIFI, operator network, NFC (Near Field Communication) or other technologies.
  • the computer program when executed by a processor, implements a network communication method.
  • the display screen of the computer equipment may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment may be a touch layer covered on the display screen, or a button, a trackball or a touchpad set on the shell of the computer equipment , or an external keyboard, trackpad, or mouse.
  • FIGS. 7 and 8 are only block diagrams of partial structures related to the solution of the present application, and do not constitute a limitation on the computer equipment to which the solution of the present application is applied.
  • a device may include more or fewer components than shown in the figures, or combine certain components, or have a different arrangement of components.
  • a computer device comprising a memory and one or more processors, the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, one or more processors execute the above-mentioned lane change prediction regression model training method A step of.
  • a computer device includes a memory and one or more processors, where computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processors, cause the one or more processors to perform the steps of the above-mentioned lane change prediction method.
  • One or more computer storage media storing computer-readable instructions, when executed by one or more processors, cause the one or more processors to perform the steps of the above-mentioned method for training a regression model for lane change prediction.
  • One or more computer storage media storing computer-readable instructions, which when executed by one or more processors, cause the one or more processors to perform the steps of the above-mentioned lane change prediction method.
  • the computer storage medium is a readable storage medium, and the readable storage medium may be non-volatile or volatile.
  • Nonvolatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in various forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Road (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Image Analysis (AREA)

Abstract

一种变道预测回归模型训练方法，包括：获取第一样本运动特征，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的；获取所述第一样本运动特征对应的标准变道概率，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；将所述第一样本运动特征输入到待训练的变道预测回归模型中，得到所述第一样本运动特征对应的预测变道概率；根据所述预测变道概率与所述标准变道概率之间的概率差异，得到模型损失值；利用所述模型损失值调整所述变道预测回归模型中的模型参数，得到所述已训练的变道预测回归模型。

Description

变道预测回归模型训练方法、变道预测方法和装置 技术领域
本申请涉及一种变道预测回归模型训练方法、变道预测方法、装置、计算机设备和存储介质。
背景技术
随着人工智能的发展，出现了自动驾驶汽车，自动驾驶汽车是一种通过计算机***实现无人驾驶的智能汽车，其依靠人工智能、视觉计算、雷达、监控装置和全球定位***协同合作，使得计算机***在没有人类的主动操作下，自动并且安全地控制汽车行驶。在自动驾驶汽车的行驶过程中，需要对行驶途中或者行驶路径附近的运动物体的运动情况进行检测，以躲避运动物体，从而保证自动驾驶汽车的安全性。
然而，发明人意识到，目前的用于识别运动物体的运动状态的方式存在不准确的情况，导致自动驾驶汽车的避障能力低，从而使得自动驾驶车辆的安全性低。
发明内容
根据本申请公开的各种实施例，提供一种变道预测回归模型训练方法、变道预测方法、装置、计算机设备和存储介质。
一种变道预测回归模型训练方法，包括：
获取第一样本运动特征，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的；
获取所述第一样本运动特征对应的标准变道概率，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；
将所述第一样本运动特征输入到待训练的变道预测回归模型中，得到所述第一样本运动特征对应的预测变道概率；
根据所述预测变道概率与所述标准变道概率之间的概率差异，得到模型损失值；及
利用所述模型损失值调整所述变道预测回归模型中的模型参数，得到所述已训练的变道预测回归模型，以根据所述已训练的变道预测回归模型进行变道预测。
一种变道预测回归模型训练装置，包括：
第一样本运动特征获取模块，用于获取第一样本运动特征，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的；
标准变道概率获取模块，用于获取所述第一样本运动特征对应的标准变道概率，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间 的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；
预测变道概率得到模块，用于将所述第一样本运动特征输入到待训练的变道预测回归模型中，得到所述第一样本运动特征对应的预测变道概率；
模型损失值得到模块，用于根据所述预测变道概率与所述标准变道概率之间的概率差异，得到模型损失值；及
已训练的变道预测回归模型得到模块，用于利用所述模型损失值调整所述变道预测回归模型中的模型参数，得到所述已训练的变道预测回归模型，以根据所述已训练的变道预测回归模型进行变道预测。
一种计算机设备，包括存储器和一个或多个处理器，所述存储器中储存有计算机可读指令，所述计算机可读指令被所述处理器执行时，使得所述一个或多个处理器执行以下步骤：
获取第一样本运动特征，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的；
获取所述第一样本运动特征对应的标准变道概率，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；
将所述第一样本运动特征输入到待训练的变道预测回归模型中，得到所述第一样本运动特征对应的预测变道概率；
根据所述预测变道概率与所述标准变道概率之间的概率差异，得到模型损失值；及
利用所述模型损失值调整所述变道预测回归模型中的模型参数，得到所述已训练的变道预测回归模型，以根据所述已训练的变道预测回归模型进行变道预测。
一个或多个存储有计算机可读指令的计算机存储介质，计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器执行以下步骤：
获取第一样本运动特征，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的；
获取所述第一样本运动特征对应的标准变道概率，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；
将所述第一样本运动特征输入到待训练的变道预测回归模型中，得到所述第一样本运动特征对应的预测变道概率；
根据所述预测变道概率与所述标准变道概率之间的概率差异，得到模型损失值；及
利用所述模型损失值调整所述变道预测回归模型中的模型参数，得到所述已训练的变道预测回归模型，以根据所述已训练的变道预测回归模型进行变道预测。
一种变道预测方法，包括：
获取目标运动对象对应的目标运动特征；
将所述目标运动特征输入到已训练的变道预测回归模型中，得到所述目标运动对象对应的目标变道概率；所述已训练的变道预测回归模型根据第一样本运动特征以及所述第一样本运动特征对应的标准变道概率训练得到，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；及
根据所述目标变道概率确定所述目标运动对象对应的目标变道结果。
一种变道预测装置，包括：
目标运动特征获取模块，用于获取目标运动对象对应的目标运动特征；
目标变道概率得到模块，用于将所述目标运动特征输入到已训练的变道预测回归模型中，得到所述目标运动对象对应的目标变道概率；所述已训练的变道预测回归模型根据第一样本运动特征以及所述第一样本运动特征对应的标准变道概率训练得到，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；及
目标变道结果确定模块，用于根据所述目标变道概率确定所述目标运动对象对应的目标变道结果。
一种计算机设备，包括存储器和一个或多个处理器，所述存储器中储存有计算机可读指令，所述计算机可读指令被所述处理器执行时，使得所述一个或多个处理器执行以下步骤：
获取目标运动对象对应的目标运动特征；
将所述目标运动特征输入到已训练的变道预测回归模型中，得到所述目标运动对象对应的目标变道概率；所述已训练的变道预测回归模型根据第一样本运动特征以及所述第一样本运动特征对应的标准变道概率训练得到，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；及
根据所述目标变道概率确定所述目标运动对象对应的目标变道结果。
一个或多个存储有计算机可读指令的计算机存储介质，计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器执行以下步骤：
获取目标运动对象对应的目标运动特征；
将所述目标运动特征输入到已训练的变道预测回归模型中，得到所述目标运动对象对应的目标变道概率；所述已训练的变道预测回归模型根据第一样本运动特征以及所述第一样本运动特征对应的标准变道概率训练得到，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，所述标准变道概率根据所述第一样本运动特 征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；及
根据所述目标变道概率确定所述目标运动对象对应的目标变道结果。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其它特征和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1为根据一个或多个实施例中变道预测回归模型训练方法的应用场景图；
图2为根据一个或多个实施例中变道预测回归模型训练方法的流程示意图；
图3为根据一个或多个实施例中得到目标时间间隔阈值以及目标变道概率生成因子的步骤的流程示意图；
图4为根据一个或多个实施例中变道预测方法的流程示意图；
图5为根据一个或多个实施例中变道预测回归模型训练装置的框图；
图6为根据一个或多个实施例中变道预测装置的框图；
图7为根据一个或多个实施例中计算机设备的框图；
图8为根据一个或多个实施例中计算机设备的框图。
具体实施方式
为了使本申请的技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
本申请提供的变道预测回归模型训练方法，可以应用于如图1所示的应用环境中。终端102与服务器104通过网络进行通信。服务器104获取第一样本运动特征，第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，获取第一样本运动特征对应的标准变道概率，标准变道概率根据第一样本运动特征对应的对象变道时间与样本采集时间之间的变道时间间隔确定，标准变道概率与变道时间间隔成负相关关系，将第一样本运动特征输入到待训练的变道预测回归模型中，得到第一样本运动特征对应的预测变道概率，根据预测变道概率与标准变道概率之间的概率差异，得到模型损失值，利用模型损失值调整变道预测回归模型中的模型参数，得到已训练的变道预测回归模型，以根据已训练的变道预测回归模型进行变道预测。服务器104可以利用已训练的变道预测回归模型进行变道预测，根据变道预测的结果控制运动，服务器104还可以将已训练的变道预测回归模型传输至终端102，终端102可以利用已训练的变道预测回归模型进行变道预测， 服务器或者终端根据变道预测的结果控制终端进行运动。终端102可以但不限于是自动驾驶汽车和移动机器人，服务器104可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一些实施例中，如图2所示，提供了一种变道预测回归模型训练方法，以该方法应用于图1中的服务器104为例进行说明，包括以下步骤：
S202，获取第一样本运动特征，第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的。
具体地，运动对象指的是处于运动状态的对象。对象可以是有生命的物体，例如可以包括人或动物中的至少一种，也可以是无生命的物体，例如可以是车辆、飞行器或自行车中的至少一种。样本运动对象可以是任意的运动对象，也可以是指定的运动对象。
运动数据可以是样本采集时间采集到的与样本运动对象的运动相关的数据，可以包括样本运动对象运动过程中点云采集设备采集到的点云(point cloud)数据，也可以包括图像采集设备采集到的图像数据中的至少一种。点云指的是三维坐标系中的三维数据点组成的集合，例如可以是物体的表面在三维坐标系中对应的各个三维数据点组成的集合，点云可以表示一个物体的外表面形状。三维数据点指的是三维空间中的点。三维数据点还可以包括RGB颜色、灰度值或时间中的至少一种。点云可以是通过激光雷达扫描得到的。点云采集设备可以是任意的可以采集点云数据的设备，可以但不限于是激光雷达，例如可以是自动驾驶车辆顶部设置的激光雷达。图像采集设备可以是任意的可以采集图像数据的设备，可以但不限于是相机。激光雷达是一种有源的传感器，通过发射激光束，将激光束打到物体表面之后，激光束被反弹，收集反弹的激光信号得到物体的点云。运动数据可以是预先存储在服务器中的。运动数据可以是样本运动对象中安装的点云采集设备或图像采集设备采集得到的，也可以是样本运动对象周围环境中的点云采集设备或图像采集设备采集得到的，例如可以是与样本运动对象之间的距离小于距离阈值的对象中安装的设备采集得到的。
运动特征指的是与运动相关的特征，根据运动数据计算得到。运动特征可以是根据从运动数据中提取到的运动特征相关数据计算得到的，例如运动特征可以为运动特征相关数据，也可以是根据不同时刻的运动数据对应的运动特征相关数据计算得到的。例如，服务器可以从点云数据中选取至少两个的点云帧，从选取的各个点云帧中提取运动特征相关数据，对运动特征相关数据进行数据拟合，得到拟合结果，将拟合结果作为运动特征，例如可以计算运动特征相关数据的统计值，得到运动特征。统计值可以包括均值或方差中的至少一种。运动特征相关数据可以是通过神经网络模型从运动数据中提取到的。例如，服务器可以将运动数据输入到已训练的运动特征相关数据识别模型中，运动特征相关数据识别模型可以对运动数据进行处理，例如进行卷积处理，得到运动特征相关数据。
运动特征相关数据可以包括运动对象的位置、速度、加速度、路程、运动方向或相对距离中的至少一种，例如可以包括运动对象在世界坐标系中的位置、速度、加速度或者 运动方向中的至少一种。世界坐标系可以为三维的坐标系，世界坐标系中的位置可以用(x,y,z)表示，x、y以及z为互相垂直并相交的坐标轴。速度例如可以包括运动对象相对于道路中心线的水平速度或垂直速度中的至少一种，水平速度指的是与道路中心线平行的方向上的速度，垂直速度指的是与道路中心线垂直的方向上的速度。运动方向例如可以是车辆朝向。相对距离可以是样本运动对象与参照物之间的距离，参照物可以是处于静止状态的物体，例如可以是道路边界或道路上设置的物体中的至少一种，道路上设置的物体例如可以是树木。参照物也可以是虚拟的事物，例如可以是航线，航线指的是飞行器飞行的路线，也可以是三维坐标系中的位置，例如三维坐标系的原点。道路边界也可以称为车道边界。即相对距离可以是运动对象与道路边界的距离，运动对象与道路边界的距离可以称为道路边界距离。道路边界距离可以包括道路左边界距离或道路右边界距离中的至少一个。道路左边界距离指的是样本运动对象与道路左边界的距离，道路右边界距离指的是样本运动对象与道路右边界的距离。
运动特征还可以是根据运动特征相关数据的排序结果确定的，可以将运动特征相关数据的排序结果作为运动特征，例如可以将不同时刻的相对距离按照时间先后顺序进行排序，得到距离排序结果，作为运动特征，或者对运动特征相关数据的排序结果进行计算得到运动特征，例如可以对距离排序结果中的数据进行归一化处理，得到运动特征，或者对距离排序结果中的数据进行统计计算或数据拟合，得到运动特征，或者对运动特征相关数据的排序结果进行拼接得到拼接结果，作为运动特征，例如可以将距离排序结果与速度排序结果进行拼接得到运动特征，速度排序结果指的是将不同时刻的速度按照时间先后顺序进行排序得到的排序结果，例如可以是将不同时刻的水平速度按照时间先后顺序进行排序得到的排序结果，例如距离排序结果为(s1,s2,s3)，速度排序结果为(v1,v2,v3)，则距离排序结果与速度排序结果拼接后的结果可以表示为(s1,s2,s3，v1,v2,v3)。当然，也可以对拼接结果中的数据进行统计计算或者归一化处理，得到运动特征。距离排序结果中的相对距离的数量与速度排序结果中速度的数量可以是一致的，也可以是不一致的。距离排序结果中的相对距离对应的点云帧，与速度排序结果中速度对应的点云帧可以是一致的，也可以是不一致的。
样本运动特征指的是样本运动对象的运动特征，第一样本运动特征可以是任意的样本运动特征，也可以是指定的样本运动特征，用于对变道预测回归模型进行训练，得到已训练的变道预测回归模型。变道预测回归模型是用于预测发生变道的概率的回归模型。变道预测回归模型的输入可以是运动特征，输出可以是发生变道的概率。变道预测回归模型可以是现有的回归模型，也可以是自定义的回归模型，例如可以是一个四层的神经网络。回归模型也可以称为回归器和回归神经网络。样本采集时间指的是采集得到运动数据的时间，可以是时间点，也可以是时间段，例如可以是(t1，t2)，其中t1表示样本采集时间的起始时间，t2表示样本采集时间的终止时间，时间段的时长例如可以是5分钟。
在一些实施例中，服务器可以从点云采集设备获取样本运动对象在运动过程中的运 动点云，从获取运动点云包括的点云帧中获取第一数量的点云帧，确定第一数量的点云帧中各个点云帧中样本运动对象的运动特征相关数据，按照点云帧的时间先后顺序，对各个点云帧分别对应的运动特征相关数据进行排序，得到排序结果，根据排序结果得到样本运动特征，例如计算排序结果中数据的统计值，得到样本运动特征，或者对排序结果中的数据进行数据拟合，根据数据拟合的结果得到样本运动特征，例如可以将数据拟合的结果作为样本运动特征，或者对数据拟合的结果进行归一化处理得到样本运动特征。第一数量的点云帧中的点云帧之间可以是连续的，也可以是不连续的点云帧。第一数量可以是预先设置的，也可以根据需要设置，例如可以是10。
S204，获取第一样本运动特征对应的标准变道概率，标准变道概率根据第一样本运动特征对应的对象变道时间与样本采集时间之间的变道时间间隔确定，标准变道概率与变道时间间隔成负相关关系。
具体地，对象变道时间指的是样本运动对象发生变道的时间。第一样本运动特征对应的对象变道时间指的是样本采集时间之后样本运动对象第一次发生变道的时间。例如，样本采集时间(t1，t2)之后样本运动对象在t2+a1时刻发生了变道，则t2+a1为对象变道时间。变道时间间隔指的是第一样本运动特征对应的对象变道时间与样本采集时间之间的时间间隔。第一样本运动特征对应的标准变道概率指的是第一样本运动特征的真实的变道概率。对于任意的变道时间间隔，标准变道概率可以与变道时间间隔成负相关关系。当然，也可以是当变道时间间隔满足时间间隔条件时，标准变道概率与变道时间间隔成负相关关系。时间间隔条件可以包括变道时间间隔小于或等于目标时间间隔阈值。目标时间间隔阈值可以是预先设置的，也可以是通过对变道预测回归模型进行训练确定的，例如可以是通过交叉验证确定的。当标准变道概率与变道时间间隔成负相关关系时，标准变道概率可以是根据变道时间间隔的平方值计算得到的，例如可以是根据变道时间间隔的平方值与特定数值的比值得到的，特定数值可以是预先设置的，也可以是通过对变道预测回归模型进行训练确定的。
在一些实施例中，当变道时间间隔大于目标时间间隔阈值时，服务器可以获取固定变道概率，作为标准变道概率。固定变道概率可以是根据需要设置的，也可以是预先设置的，例如可以为0。
S206，将第一样本运动特征输入到待训练的变道预测回归模型中，得到第一样本运动特征对应的预测变道概率。
具体地，待训练的变道预测回归模型指的是需要进行训练的变道预测回归模型。服务器可以将第一样本运动特征作为待训练的变道预测回归模型的输入，待训练的变道预测回归模型可以对第一样本运动特征进行计算，例如进行卷积计算，得到第一样本运动特征对应的预测变道概率。预测变道概率是变道预测回归模型输出的变道概率。
S208，根据预测变道概率与标准变道概率之间的概率差异，得到模型损失值。
具体地，概率差异指的是预测变道概率与标准变道概率之间的差异。模型损失值与 概率差异成正相关关系。模型损失值例如可以是概率差异的平方或概率差异的平方的倍数中的任意一种，倍数例如为二分之一倍。
在一些实施例中，模型损失值可以根据概率差异与差异阈值之间的大小关系确定。例如当概率差异小于差异阈值时，可以根据概率差异的平方确定模型损失值，模型损失值与概率差异的平方成正相关关系。当概率差异大于或者等于差异阈值时，可以根据概率差异以及差异阈值计算得到模型损失值。差异阈值可以是预先设置的，也可以是根据需要设置的。
S210，利用模型损失值调整变道预测回归模型中的模型参数，得到已训练的变道预测回归模型，以根据已训练的变道预测回归模型进行变道预测。
具体地，模型参数指的是变道预测回归模型内部的变量参数，对于神经网络模型，也可以称为神经网络权重(weight)。已训练的变道预测回归模型可以是经过一次或者多次训练得到的。例如，服务器可以朝着损失值变小的方向调整变道预测回归模型中的模型参数，可以经过多次迭代训练，得到已训练的变道预测回归模型。
在一些实施例中，服务器可以根据模型损失值进行反向传播，并在反向传播的过程中，沿梯度下降方向更新变道预测回归模型的模型参数，得到已训练的变道预测回归模型。其中，反向是指参数的更新与变道的预测的方向是相反的，由于参数的更新是反向传播的，因此可以根据模型损失值得到下降梯度，从变道预测回归模型的最后一层开始，根据下降梯度开始进行模型参数的梯度更新，直至到达变道预测回归模型的第一层。梯度下降方法可以为随机梯度下降法或批量梯度下降中的任意一种。可以理解，模型的训练可以是迭代多次的，即已训练的变道预测回归模型可以是迭代训练得到的，当满足模型收敛条件时再停止训练，模型收敛条件可以是模型损失值的小于预设损失值，也可以是模型参数的变化小于预设参数变化值。
在一些实施例中，服务器可以将已训练的变道预测回归模型传输至终端，终端利用已训练的变道预测回归模型对周围环境中的运动对象进行变道预测，从而控制终端的运动，实现避障。
上述变道预测回归模型训练方法中，获取第一样本运动特征，第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，获取第一样本运动特征对应的标准变道概率，将第一样本运动特征输入到待训练的变道预测回归模型中，得到第一样本运动特征对应的预测变道概率，根据预测变道概率与标准变道概率之间的概率差异，得到模型损失值，利用模型损失值调整变道预测回归模型中的模型参数，得到已训练的变道预测回归模型，以根据已训练的变道预测回归模型进行变道预测，由于标准变道概率根据第一样本运动特征对应的对象变道时间与样本采集时间之间的变道时间间隔确定，标准变道概率与变道时间间隔成负相关关系，从而提高了已训练的变道预测回归模型的准确度，提高了变道预测的准确度。
在一些实施例中，获取第一样本运动特征对应的标准变道概率包括：当变道时间间隔 小于或等于目标时间间隔阈值时，获取目标变道概率生成因子；及，根据变道时间间隔以及目标变道概率生成因子进行计算，得到第一样本运动特征对应的标准变道概率。
具体地，时间间隔阈值可以为一段时长，例如可以是10秒。目标时间间隔阈值可以是从候选时间间隔阈值中选取得到的。候选时间间隔阈值可以是服务器中预先存储的时间间隔阈值，也可以是服务器获取或者生成的时间间隔阈值。变道概率生成因子用于生成变道概率。变道概率指的是运动对象发生变道的概率。目标变道概率生成因子可以是从候选变道概率生成因子中选取得到的。候选变道概率生成因子可以是服务器中预先存储的变道概率生成因子，也可以是服务器获取或者生成的变道概率生成因子。目标变道概率生成因子可以是通过对变道预测回归模型进行训练确定的。例如，服务器可以根据候选变道概率生成因子以及变道时间间隔生成候选变道概率，将候选变道概率作为第二样本运动特征的概率标签，利用第二样本运动特征以及对应的概率标签对变道预测回归模型进行训练，根据训练后的变道预测回归模型的准确度确定目标变道概率生成因子。第二样本运动特征可以是任意的样本运动特征，也可以是指定的样本运动特征，例如可以是与第一样本运动特征不同的样本运动特征。概率标签也可以称为回归真值。
在一些实施例中，服务器可以利用变道概率计算公式，对变道时间间隔以及目标变道概率生成因子进行计算，得到第一样本运动特征对应的标准变道概率。变道概率计算公式中，当自变量的取值大于或等于时，因变量随着自变量的增大而减小，并且因变量的取值大于0且小于1。服务器可以将变道概率计算公式中的自变量替换为变道时间间隔，将变道概率计算公式中的常量替换为目标变道概率生成因子，计算得到第一样本运动特征对应的标准变道概率。变道概率计算公式例如可以是高斯函数。服务器可以利用高斯函数对变道时间间隔以及目标变道概率生成因子进行计算，得到第一样本运动特征对应的标准变道概率。例如，服务器可以将高斯函数的自变量替换为变道时间间隔，将高斯函数中的方差替换为目标变道概率生成因子，高斯函数的均值设置为0，计算得到第一样本运动特征对应的标准变道概率。
上述实施例中，当变道时间间隔小于或等于目标时间间隔阈值时，获取目标变道概率生成因子，根据变道时间间隔以及目标变道概率生成因子进行计算，得到第一样本运动特征对应的标准变道概率，从而可以保证标准变道概率对应的变道时间间隔不至于过大，保证变道时间间隔的有效性，提高标准变道概率的准确度。
在一些实施例中，根据变道时间间隔以及目标变道概率生成因子进行计算，得到第一样本运动特征对应的标准变道概率包括：对变道时间间隔进行平方运算，得到间隔平方值；计算间隔平方值与目标变道概率生成因子的比值，得到目标比值；及，将第一数值作为底数，将目标比值的负数作为指数进行指数计算，得到第一样本运动特征对应的标准变道概率。
具体地，间隔平方值为变道时间间隔的平方，例如变道时间间隔为d，则间隔平方值为d 2。目标比值指的是间隔平方值与目标变道概率生成因子的比值，例如可以为d 2/C，C 表示目标变道概率生成因子。第一数值可以为任意的大于1的正数，例如可以为自然对数的底数e≈2.71828。
在一些实施例中，服务器可以将第一数值作为底数，根据目标比值确定指数进行指数计算得到标准变道概率。例如，服务器可以将第一数值作为底数，将目标比值的负数作为指数进行指数计算，得到第一样本运动特征对应的标准变道概率。服务器可以将第一数值作为底数，将目标比值的负数与第二数值的比值作为指数进行指数计算，得到标准变道概率。第二数值可以为任意的正数，例如可以是2。标准变道概率例如可以表示为公式(1)，其中c 2表示目标变道概率生成因子。f(d)表示标准变道概率。
Figure PCTCN2020128239-appb-000001
上述实施例中，对变道时间间隔进行平方运算，得到间隔平方值，计算间隔平方值与目标变道概率生成因子的比值，得到目标比值，将第一数值作为底数，将目标比值的负数作为指数进行指数计算，得到第一样本运动特征对应的标准变道概率，从而可以保证标准变道概率的取值在0和1之间，提高了标准变道概率的准确度。
在一些实施例中，如图3所示，得到目标时间间隔阈值以及目标变道概率生成因子的步骤包括：
S302，确定至少两个候选参数组合，候选参数组合包括候选时间间隔阈值以及候选变道概率生成因子。
具体地，服务器中可以从多个候选时间间隔阈值中任意选取一个候选时间间隔阈值，作为候选参数组合中的候选时间间隔阈值，可以从多个候选变道概率生成因子中任意选取一个候选变道概率生成因子，作为候选参数组合中的候选时间间隔阈值。若候选时间间隔阈值的数量为N，候选变道概率生成因子的数量为M，候选参数组合的数量可以为N×M。例如，候选时间间隔阈值为r1以及r2，候选变道概率生成因子为w1以及w2，则候选参数组合可以包括(r1,w1)、(r1,w2)、(r2,w1)以及(r2,w2)。
S304，获取第二样本运动特征，根据候选参数组合确定第二样本运动特征对应的变道概率，作为第二样本运动特征对应的概率标签。
具体地，第二样本运动特征可以是任意的样本运动特征，也可以是指定的样本运动特征，例如可以是与第一样本运动特征不同的样本运动特征。第二样本运动特征可以有多个。概率标签指的是第二样本运动特征的标签，概率标签也可以称为回归真值。概率标签与标准变道概率的得到方式类似，概率标签根据第二样本运动特征对应的变道时间间隔确定，当变道时间间隔小于或等于候选参数组合中的候选时间间隔阈值，根据变道时间间隔以及候选参数组合中的候选变道概率生成因子进行计算，得到第二样本运动特征对应的概率标签。当变道时间间隔大于候选参数组合中的候选时间间隔阈值，将第二样本运动特征对应的概率标签确定为固定概率，固定概率可以根据需要设置，也可以是预先设置的，例如可以是0。
在一些实施例中，服务器可以获取候选样本运动特征集合，从候选样本运动特征集合 中选取第二数量的候选样本运动特征，得到第二样本运动特征集合。将候选样本运动特征集合中第二样本运动特征集合以外的候选样本运动特征组成第三样本运动特征集合。第二数量可以根据需要设置，例如根据候选样本运动特征集合中候选样本运动特征的数量(记作特征总数)确定，第二数量例如为特征总数的五分之一。
S306，基于第二样本运动特征以及第二样本运动特征对应的概率标签，对变道预测回归模型进行训练，得到候选参数组合对应的训练后的变道预测回归模型。
具体地，服务器可以将第二样本运动特征输入到变道预测回归模型中，得到第二样本运动特征对应的预测变道概率，服务器可以根据预测变道概率与概率标签之间的概率差异，得到模型损失值，利用模型损失值调整变道预测回归模型中的模型参数，得到训练后的变道预测回归模型。例如，可以将第二样本运动特征集合中的各个第二样本运动特征输入到变道预测回归模型中，得到训练后的变道预测回归模型。
在一些实施例中，服务器可以从候选样本运动特征集合中，获取多个第二样本运动特征集合，不同的第二样本运动特征集合中的第二样本运动特征不相同。利用同一候选参数组合，确定各个第二样本运动特征集合中第二样本运动特征的概率标签。分别利用各个第二样本运动特征集合对变道预测回归模型进行训练，得到同一候选参数组合下，不同第二样本运动特征集合得到的训练后的变道预测回归模型。
S308，确定候选参数组合对应的训练后的变道预测回归模型的模型准确度，基于模型准确度从至少两个候选参数组合中选取得到满足准确度条件的目标参数组合，将目标参数组合中的候选时间间隔阈值作为目标时间间隔阈值，将目标参数组合中的候选变道概率生成因子作为目标变道概率生成因子。
具体地，候选参数组合还可以包括候选变道概率阈值。服务器可以将第三样本运动特征集合中的第三样本运动特征输入到训练后的变道预测回归模型中，得到变道预测回归模型输出的第三样本运动特征的预测变道概率，将预测变道概率与候选变道概率阈值进行对比，当预测变道概率大于候选变道概率阈值时，确定第三样本运动特征的预测变道结果为变道，当预测变道概率小于或等于候选变道概率阈值时，确定第三样本运动特征的预测变道结果为不变道。预测变道结果为预测的变道结果。服务器可以获取第三样本运动特征的标准变道结果，将预测变道结果与标准变道结果进行对比，当对比一致时，确定训练后的变道预测回归模型对第三样本运动特征的预测结果是正确的，反之预测结果是错误的。标准变道结果为第三样本运动特征的真实的变道结果。服务器可以根据各个第三样本运动特征分别对应的预测结果，确定训练后的变道预测回归模型的模型准确度，例如可以用正确的预测结果的数量除以全部预测结果的数量，得到模型准确度。准确度条件可以包括准确度大于准确度阈值或模型准确度最大中的至少一种。准确度阈值可以是根据需要设置的，也可以是预先设置的，例如可以是95％。目标参数组合为获取的至少两个候选参数组合中满足准确度条件的候选参数组合。
在一些实施例中，服务器可以确定同一候选参数组合下，不同第二样本运动特征集合 得到的训练后的变道预测回归模型分别对应的准确度，计算各个准确度的平均值，作为候选参数组合对应的训练后的变道预测回归模型的模型准确度。服务器可以确定该至少两个候选参数组合中各个候选参数组合分别对应的模型准确度，从各个候选参数组合分别对应的模型准确度中选取满足准确度条件的候选参数组合，作为目标参数组合，例如选取最大的模型准确度对应的候选参数组合，作为目标参数组合。服务器可以将目标参数组合中的候选变道概率阈值作为目标变道概率阈值。
在一些实施例中，服务器可以将目标参数组合对应的训练后的变道预测回归模型，作为待训练的变道预测回归模型。
上述实施例中，确定候选参数组合对应的训练后的变道预测回归模型的模型准确度，基于模型准确度从至少两个候选参数组合中选取得到满足准确度条件的目标参数组合，将目标参数组合中的候选时间间隔阈值作为目标时间间隔阈值，将目标参数组合中的候选变道概率生成因子作为目标变道概率生成因子，从而得到可以使得变道预测回归模型的准确度高的目标时间间隔阈值以及目标变道概率生成因子。
在一些实施例中，第一样本运动特征包括距离变化特征，确定距离变化特征的步骤包括：获取样本采集时间中，样本运动对象对应的目标距离序列，目标距离序列按照运动时间排序；根据目标距离序列中的距离的大小关系确定距离变化趋势；及，根据距离变化趋势确定样本运动对象对应的距离变化特征。
具体地，距离序列指的是相对距离组成的序列，距离序列中各个相对距离按照运动时间排序。距离序列中，运动时间越靠前，相对距离的排序越靠前，例如可以为(s1,s2,s3,s4,s5,s6)，其中s1到S6分别表示一个相对距离。运动时间指的是样本运动对象运动过程中的时间，可以是时间点，也可以是时间段。样本运动对象可以对应有多个距离序列，例如可以对应有左边界距离序列或右边界距离序列中的至少一种。左边界距离序列指的是按照时间先后顺序对道路左边界距离进行排序得到的序列，右边界距离序列指的是按照时间先后顺序对道路右边界距离进行排序得到的序列。距离序列也可以用向量表示。样本运动对象可以对应有多个距离序列，例如服务器可以根据在样本采集时间采集到的运动数据得到样本运动对象对应的多个距离序列，目标距离序列可以是从样本运动对象对应的各个距离序列中选取的距离序列。可以采用随机选取的方式选取，也可以采用预设的选取方式选取，例如可以根据距离序列对应的距离变化趋势进行选取，将满足距离变化趋势条件的距离序列作为目标距离序列。当然，可以将样本运动对象的各个距离序列分别作为目标距离序列。目标距离序列可以有多个。距离变化趋势指的是距离序列中的距离随时间变化的规律，可以包括逐渐变小的变化趋势或逐渐变大的变化趋势中的任意一种。距离变化趋势条件例如可以是逐渐变小的变化趋势。例如，当左边界距离序列的距离变化趋势为逐渐变小的变化趋势时，将左边界距离序列作为目标距离序列，当左边界距离序列的距离变化趋势为逐渐变大的变化趋势时，将右边界距离序列作为目标距离序列。
距离变化特征可以是距离变化趋势的特征，可以根据距离变化趋势计算得到，例如 可以根据距离变化趋势确定距离变化的快慢程度，得到距离变化特征。距离变化的快慢程度可以包括距离变大的快慢程度或距离变小的快慢程度中的至少一种。距离变化特征还可以根据平均距离变化程度确定，平均距离变化程度指的是单位时间内距离变化的大小，可以根据目标距离序列以及目标距离序列对应的时间序列计算得到。目标距离序列对应的时间序列包括目标距离序列中各个距离分别对应的时间，时间序列中的时间按照从前到后的顺序排序，时间越靠前，排序越靠前。时间序列例如可以是(t1,t2,t3,t4,t5,t6)。其中，t1发生在t6之前。例如可以通过目标距离序列中起始位置的距离与终止位置的距离之间的差异，得到距离差异，根据时间序列中起始位置的时间与终止位置之间的差异，得到时间差异，根据距离差异与时间差异的比值得到平均距离变化程度，平均距离变化程度例如可以为(s6-s1)÷(t6-t1)。距离变化特征可以通过向量表示，例如可以称为距离特征向量。距离变化特征还可以为目标距离序列，即可以将目标距离序列作为距离变化特征。
在一些实施例中，服务器可以利用变化特征计算方法，对目标距离序列以及对应的时间序列进行计算，得到距离变换特征。变化特征计算方法例如可以是公式(2)和(3)。其中，x表示目标距离序列中的距离，y表示目标距离序列中的距离对应的时间。a表示距离变化特征，也可以理解为相对距离随时间变化的变化率(slope)。
Figure PCTCN2020128239-appb-000002
表示y的平均值，
Figure PCTCN2020128239-appb-000003
表示x的平均值。
Figure PCTCN2020128239-appb-000004
Figure PCTCN2020128239-appb-000005
上述实施例中，获取样本采集时间中，样本运动对象对应的目标距离序列，目标距离序列按照运动时间排序，根据目标距离序列中的距离的大小关系确定距离变化趋势，根据距离变化趋势确定样本运动对象对应的距离变化特征，提高了第一样本运动特征的准确度。
在一些实施例中，如图4所示，提供了一种变道预测方法，以该方法应用于图1中的终端102为例进行说明，包括以下步骤：
S402，获取目标运动对象对应的目标运动特征。
具体地，目标运动对象可以是终端的周边环境中存在的运动对象。终端的周边环境例如可以是以终端为圆心，固定半径为半径的圆围成的区域。固定半径可以是根据需要设置的，也可以是预先设置的。目标运动特征指的是目标运动对象对应的当前时间的运动特征。当前时间为时间段，当前时间的终止时刻为当前时刻，当前时间的起始时刻为与当前时刻的时间间隔为第一时间间隔的历史时刻。第一时间间隔可以是预先设置的，也可以是根据需要确定的。
S404，将目标运动特征输入到已训练的变道预测回归模型中，得到目标运动对象对应 的目标变道概率；已训练的变道预测回归模型根据第一样本运动特征以及第一样本运动特征对应的标准变道概率训练得到，第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，标准变道概率根据第一样本运动特征对应的对象变道时间与样本采集时间之间的变道时间间隔确定，标准变道概率与变道时间间隔成负相关关系。
具体地，已训练的变道预测回归模型可以是采用上述提出的变道预测回归模型训练方法训练得到的。第一时间间隔可以与样本采集时间的时长一致。
S406，根据目标变道概率确定目标运动对象对应的目标变道结果。
具体地，目标变道结果可以包括变道或不变道中的任意一种，还可以包括预测变道时间信息。预测变道时间信息用于预测发生变道的时间，例如可以是未来的5秒内。目标变道结果为变道表示目标运动对象即将发生变道，例如未来的5秒内可能发生变道。变道可以用1表示，不变道可以用0表示。
在一些实施例中，终端可以根据目标变道结果控制终端的运动，以避免与目标运动对象发生碰撞。例如当终端前方道路上的运动对象的目标变道结果为变道时，终端可以减速行驶。
上述变道预测方法中，获取目标运动对象对应的目标运动特征，将目标运动特征输入到已训练的变道预测回归模型中，得到目标运动对象对应的目标变道概率，根据目标变道概率确定目标运动对象对应的目标变道结果，由于已训练的变道预测回归模型根据第一样本运动特征以及第一样本运动特征对应的标准变道概率训练得到，第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，标准变道概率根据第一样本运动特征对应的对象变道时间与样本采集时间之间的变道时间间隔确定，标准变道概率与变道时间间隔成负相关关系，从而提高了已训练的变道预测回归模型的准确度，提高了变道预测的准确度。
在一些实施例中，根据目标变道概率确定目标运动对象对应的目标变道结果，包括：将目标变道概率与目标变道概率阈值进行对比，得到对比结果；及，根据对比结果确定目标运动对象对应的目标变道结果。
具体地，目标变道概率阈值可以是预先设置的，也可以是通过对变道预测回归模型进行训练确定的，例如可以是0.9。对比结果可以为目标变道概率大于目标变道概率阈值、目标变道概率小于或等于目标变道概率阈值中的任意一种。当对比结果为目标变道概率大于目标变道概率阈值，终端可以确定目标变道结果为变道，反之，终端可以确定目标变道结果为不变道。
上述实施例中，将目标变道概率与目标变道概率阈值进行对比，得到对比结果，根据对比结果确定目标运动对象对应的目标变道结果，提高了目标变道结果的准确度。
应该理解的是，虽然图2-4的流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，这些步骤可以以其它的顺序执行。而且，图2-4中的至少一 部分步骤可以包括多个子步骤或者多个阶段，这些子步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，这些子步骤或者阶段的执行顺序也不必然是依次进行，而是可以与其它步骤或者其它步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。
在一些实施例中，如图5所示，提供了一种变道预测回归模型训练装置，包括：第一样本运动特征获取模块502、标准变道概率获取模块504、预测变道概率得到模块506、模型损失值得到模块508和已训练的变道预测回归模型得到模块510，其中：
第一样本运动特征获取模块502，用于获取第一样本运动特征，第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的。
标准变道概率获取模块504，用于获取第一样本运动特征对应的标准变道概率，标准变道概率根据第一样本运动特征对应的对象变道时间与样本采集时间之间的变道时间间隔确定，标准变道概率与变道时间间隔成负相关关系。
预测变道概率得到模块506，用于将第一样本运动特征输入到待训练的变道预测回归模型中，得到第一样本运动特征对应的预测变道概率。
模型损失值得到模块508，用于根据预测变道概率与标准变道概率之间的概率差异，得到模型损失值。
已训练的变道预测回归模型得到模块510，用于利用模型损失值调整变道预测回归模型中的模型参数，得到已训练的变道预测回归模型，以根据已训练的变道预测回归模型进行变道预测。
在一些实施例中，标准变道概率获取模块504包括：
目标变道概率生成因子获取单元，用于当变道时间间隔小于或等于目标时间间隔阈值时，获取目标变道概率生成因子。
标准变道概率得到单元，用于根据变道时间间隔以及目标变道概率生成因子进行计算，得到第一样本运动特征对应的标准变道概率。
在一些实施例中，标准变道概率得到单元，还用于对变道时间间隔进行平方运算，得到间隔平方值；计算间隔平方值与目标变道概率生成因子的比值，得到目标比值；及，将第一数值作为底数，将目标比值的负数作为指数进行指数计算，得到第一样本运动特征对应的标准变道概率。
在一些实施例中，变道预测回归模型训练装置还包括目标时间间隔阈值得到模块，目标时间间隔阈值得到模块包括：
候选参数组合确定单元，用于确定至少两个候选参数组合，候选参数组合包括候选时间间隔阈值以及候选变道概率生成因子。
概率标签得到单元，用于获取第二样本运动特征，根据候选参数组合确定第二样本运动特征对应的变道概率，作为第二样本运动特征对应的概率标签。
训练后的变道预测回归模型得到单元，用于基于第二样本运动特征以及第二样本运动 特征对应的概率标签，对变道预测回归模型进行训练，得到候选参数组合对应的训练后的变道预测回归模型。
目标时间间隔阈值得到单元，用于确定候选参数组合对应的训练后的变道预测回归模型的模型准确度，基于模型准确度从至少两个候选参数组合中选取得到满足准确度条件的目标参数组合，将目标参数组合中的候选时间间隔阈值作为目标时间间隔阈值，将目标参数组合中的候选变道概率生成因子作为目标变道概率生成因子。
在一些实施例中，第一样本运动特征包括距离变化特征，变道预测回归模型训练装置还包括距离变化特征确定模块，距离变化特征确定模块包括：
目标距离序列获取单元，用于获取样本采集时间中，样本运动对象对应的目标距离序列，目标距离序列按照运动时间排序。
距离变化趋势确定单元，用于根据目标距离序列中的距离的大小关系确定距离变化趋势。
距离变化特征确定单元，用于根据距离变化趋势确定样本运动对象对应的距离变化特征。
关于变道预测回归模型训练装置的具体限定可以参见上文中对于变道预测回归模型训练方法的限定，在此不再赘述。上述变道预测回归模型训练装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一些实施例中，如图6所示，提供了一种变道预测装置，包括：目标运动特征获取模块602、目标变道概率得到模块604和目标变道结果确定模块606，其中：
目标运动特征获取模块602，用于获取目标运动对象对应的目标运动特征。
目标变道概率得到模块604，用于将目标运动特征输入到已训练的变道预测回归模型中，得到目标运动对象对应的目标变道概率；已训练的变道预测回归模型根据第一样本运动特征以及第一样本运动特征对应的标准变道概率训练得到，第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，标准变道概率根据第一样本运动特征对应的对象变道时间与样本采集时间之间的变道时间间隔确定，标准变道概率与变道时间间隔成负相关关系。
目标变道结果确定模块606，用于根据目标变道概率确定目标运动对象对应的目标变道结果。
在一些实施例中，目标变道结果确定模块606，包括：
对比结果得到单元，用于将目标变道概率与目标变道概率阈值进行对比，得到对比结果。
目标变道结果得到单元，用于根据对比结果确定目标运动对象对应的目标变道结果。
关于变道预测装置的具体限定可以参见上文中对于变道预测方法的限定，在此不再赘述。上述变道预测装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一些实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图7所示。该计算机设备包括通过***总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作***、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作***和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储样本运动特征、标准变道概率、预测变道概率、模型损失值以及变道预测回归模型等数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种变道预测回归模型训练方法。
在一些实施例中，提供了一种计算机设备，该计算机设备可以是终端，其内部结构图可以如图8所示。该计算机设备包括通过***总线连接的处理器、存储器、通信接口、显示屏和输入装置。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作***和计算机程序。该内存储器为非易失性存储介质中的操作***和计算机程序的运行提供环境。该计算机设备的通信接口用于与外部的终端进行有线或无线方式的通信，无线方式可通过WIFI、运营商网络、NFC(近场通信)或其他技术实现。该计算机程序被处理器执行时以实现一种网络通信方法。该计算机设备的显示屏可以是液晶显示屏或者电子墨水显示屏，该计算机设备的输入装置可以是显示屏上覆盖的触摸层，也可以是计算机设备外壳上设置的按键、轨迹球或触控板，还可以是外接的键盘、触控板或鼠标等。
本领域技术人员可以理解，图7和8中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
一种计算机设备，包括存储器和一个或多个处理器，存储器中储存有计算机可读指令，计算机可读指令被处理器执行时，使得一个或多个处理器执行上述变道预测回归模型训练方法的步骤。
一种计算机设备，包括存储器和一个或多个处理器，存储器中储存有计算机可读指令，计算机可读指令被处理器执行时，使得一个或多个处理器执行上述变道预测方法的步骤。
一个或多个存储有计算机可读指令的计算机存储介质，计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器执行上述变道预测回归模型训练方法的步骤。
一个或多个存储有计算机可读指令的计算机存储介质，计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器执行上述变道预测方法的步骤。
其中，该计算机存储介质为可读存储介质，可读存储介质可以是非易失性，也可以是易失性的。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请专利的保护范围应以所附权利要求为准。

Claims (11)

  1. 一种变道预测回归模型训练方法，包括：
    获取第一样本运动特征，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的；
    获取所述第一样本运动特征对应的标准变道概率，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；
    将所述第一样本运动特征输入到待训练的变道预测回归模型中，得到所述第一样本运动特征对应的预测变道概率；
    根据所述预测变道概率与所述标准变道概率之间的概率差异，得到模型损失值；及
    利用所述模型损失值调整所述变道预测回归模型中的模型参数，得到所述已训练的变道预测回归模型，以根据所述已训练的变道预测回归模型进行变道预测。
  2. 根据权利要求1所述的方法，其中，所述获取所述第一样本运动特征对应的标准变道概率包括：
    当所述变道时间间隔小于或等于目标时间间隔阈值时，获取目标变道概率生成因子；及
    根据所述变道时间间隔以及所述目标变道概率生成因子进行计算，得到所述第一样本运动特征对应的标准变道概率。
  3. 根据权利要求2所述的方法，其中，所述根据所述变道时间间隔以及所述目标变道概率生成因子进行计算，得到所述第一样本运动特征对应的标准变道概率包括：
    对所述变道时间间隔进行平方运算，得到间隔平方值；
    计算所述间隔平方值与所述目标变道概率生成因子的比值，得到目标比值；及
    将第一数值作为底数，将所述目标比值的负数作为指数进行指数计算，得到所述第一样本运动特征对应的标准变道概率。
  4. 根据权利要求3所述的方法，其中，得到所述目标时间间隔阈值以及所述目标变道概率生成因子的步骤包括：
    确定至少两个候选参数组合，所述候选参数组合包括候选时间间隔阈值以及候选变道概率生成因子；
    获取第二样本运动特征，根据所述候选参数组合确定所述第二样本运动特征对应的变道概率，作为所述第二样本运动特征对应的概率标签；
    基于所述第二样本运动特征以及所述第二样本运动特征对应的概率标签，对所述变道预测回归模型进行训练，得到所述候选参数组合对应的训练后的变道预测回归模型；及
    确定所述候选参数组合对应的训练后的变道预测回归模型的模型准确度，基于所述模型准确度从所述至少两个候选参数组合中选取得到满足准确度条件的目标参数组合，将所述目标参数组合中的候选时间间隔阈值作为所述目标时间间隔阈值，将所述目标参数组合 中的候选变道概率生成因子作为所述目标变道概率生成因子。
  5. 根据权利要求1所述的方法，其中，所述第一样本运动特征包括距离变化特征，确定所述距离变化特征的步骤包括：
    获取所述样本采集时间中，所述样本运动对象对应的目标距离序列，所述目标距离序列按照运动时间排序；
    根据所述目标距离序列中的距离的大小关系确定距离变化趋势；及
    根据所述距离变化趋势确定所述样本运动对象对应的距离变化特征。
  6. 一种变道预测方法，包括：
    获取目标运动对象对应的目标运动特征；
    将所述目标运动特征输入到已训练的变道预测回归模型中，得到所述目标运动对象对应的目标变道概率；所述已训练的变道预测回归模型根据第一样本运动特征以及所述第一样本运动特征对应的标准变道概率训练得到，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；及
    根据所述目标变道概率确定所述目标运动对象对应的目标变道结果。
  7. 根据权利要求6所述的方法，其中，所述根据所述目标变道概率确定所述目标运动对象对应的目标变道结果，包括：
    将所述目标变道概率与目标变道概率阈值进行对比，得到对比结果；及
    根据所述对比结果确定所述目标运动对象对应的目标变道结果。
  8. 一种变道预测回归模型训练装置，包括：
    第一样本运动特征获取模块，用于获取第一样本运动特征，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的；
    标准变道概率获取模块，用于获取所述第一样本运动特征对应的标准变道概率，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；
    预测变道概率得到模块，用于将所述第一样本运动特征输入到待训练的变道预测回归模型中，得到所述第一样本运动特征对应的预测变道概率；
    模型损失值得到模块，用于根据所述预测变道概率与所述标准变道概率之间的概率差异，得到模型损失值；及
    已训练的变道预测回归模型得到模块，用于利用所述模型损失值调整所述变道预测回归模型中的模型参数，得到所述已训练的变道预测回归模型，以根据所述已训练的变道预测回归模型进行变道预测。
  9. 一种变道预测装置，包括：
    目标运动特征获取模块，用于获取目标运动对象对应的目标运动特征；
    目标变道概率得到模块，用于将所述目标运动特征输入到已训练的变道预测回归模型中，得到所述目标运动对象对应的目标变道概率；所述已训练的变道预测回归模型根据第一样本运动特征以及所述第一样本运动特征对应的标准变道概率训练得到，所述第一样本运动特征是根据在样本采集时间采集到的样本运动对象的运动数据得到的，所述标准变道概率根据所述第一样本运动特征对应的对象变道时间与所述样本采集时间之间的变道时间间隔确定，所述标准变道概率与所述变道时间间隔成负相关关系；及
    目标变道结果确定模块，用于根据所述目标变道概率确定所述目标运动对象对应的目标变道结果。
  10. 一种计算机设备，包括存储器及一个或多个处理器，所述存储器中储存有计算机可读指令，所述计算机可读指令被所述一个或多个处理器执行时，使得所述一个或多个处理器执行权利要求1至5或6至7中任一项所述方法的步骤。
  11. 一种计算机设备，包括存储器及一个或多个处理器，所述存储器中储存有计算机可读指令，所述计算机可读指令被所述一个或多个处理器执行时，使得所述一个或多个处理器执行权利要求1至5或6至7中任一项所述方法的步骤。
PCT/CN2020/128239 2020-11-12 2020-11-12 变道预测回归模型训练方法、变道预测方法和装置 WO2022099526A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202080092993.3A CN114945961B (zh) 2020-11-12 2020-11-12 变道预测回归模型训练方法、变道预测方法和装置
PCT/CN2020/128239 WO2022099526A1 (zh) 2020-11-12 2020-11-12 变道预测回归模型训练方法、变道预测方法和装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/128239 WO2022099526A1 (zh) 2020-11-12 2020-11-12 变道预测回归模型训练方法、变道预测方法和装置

Publications (1)

Publication Number Publication Date
WO2022099526A1 true WO2022099526A1 (zh) 2022-05-19

Family

ID=81601930

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/128239 WO2022099526A1 (zh) 2020-11-12 2020-11-12 变道预测回归模型训练方法、变道预测方法和装置

Country Status (2)

Country Link
CN (1) CN114945961B (zh)
WO (1) WO2022099526A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116729381A (zh) * 2023-06-02 2023-09-12 吉林省艾优数字科技有限公司 一种无人驾驶方法及***

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9672734B1 (en) * 2016-04-08 2017-06-06 Sivalogeswaran Ratnasingam Traffic aware lane determination for human driver and autonomous vehicle driving system
CN107919027A (zh) * 2017-10-24 2018-04-17 北京汽车集团有限公司 预测车辆变道的方法、装置和***
CN110304064A (zh) * 2019-07-15 2019-10-08 广州小鹏汽车科技有限公司 一种车辆变道的控制方法及车辆控制***、车辆
CN111046897A (zh) * 2018-10-11 2020-04-21 顾泽苍 一种跨越不同空间的模糊事件概率测度的定义方法

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2010033352A (ja) * 2008-07-29 2010-02-12 Toyota Central R&D Labs Inc 車線変更警報装置及びプログラム
DE102015217498A1 (de) * 2015-09-14 2017-03-16 Volkswagen Aktiengesellschaft Verfahren und Vorrichtung in einem Kraftfahrzeug zum automatisierten Fahren
CN108473140A (zh) * 2016-02-18 2018-08-31 本田技研工业株式会社 车辆控制装置、车辆控制方法及车辆控制程序
CN108032859B (zh) * 2017-12-05 2020-08-21 智车优行科技（北京）有限公司 自动变道控制方法、装置和汽车
CN111527013B (zh) * 2017-12-27 2024-02-23 宝马股份公司 车辆变道预测
EP3569460B1 (en) * 2018-04-11 2024-03-20 Hyundai Motor Company Apparatus and method for controlling driving in vehicle
CN110244732B (zh) * 2019-06-19 2020-09-29 智邮开源通信研究院(北京)有限公司 用于自动行驶的变道方法、装置、电子设备及存储介质
CN110569783B (zh) * 2019-09-05 2022-03-25 吉林大学 一种驾驶人换道意图识别方法及***
CN110705388B (zh) * 2019-09-16 2022-04-01 清华大学 一种基于预测反馈的辅助驾驶用目标车辆换道识别方法
CN111079590B (zh) * 2019-12-04 2023-05-26 东北大学 一种无人驾驶车辆的周边车辆行为预判方法
CN111798091B (zh) * 2020-05-27 2024-04-12 中汽数据有限公司 一种基于变道持续时间的高速公路变道评分模型建立方法
CN111639814A (zh) * 2020-06-02 2020-09-08 贝壳技术有限公司 预测变动行为发生概率的方法、装置、介质和电子设备
CN111806467B (zh) * 2020-07-27 2021-09-28 福州大学 一种基于车辆行驶规律的变速动态换道轨迹规划方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9672734B1 (en) * 2016-04-08 2017-06-06 Sivalogeswaran Ratnasingam Traffic aware lane determination for human driver and autonomous vehicle driving system
CN107919027A (zh) * 2017-10-24 2018-04-17 北京汽车集团有限公司 预测车辆变道的方法、装置和***
CN111046897A (zh) * 2018-10-11 2020-04-21 顾泽苍 一种跨越不同空间的模糊事件概率测度的定义方法
CN110304064A (zh) * 2019-07-15 2019-10-08 广州小鹏汽车科技有限公司 一种车辆变道的控制方法及车辆控制***、车辆

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116729381A (zh) * 2023-06-02 2023-09-12 吉林省艾优数字科技有限公司 一种无人驾驶方法及***

Also Published As

Publication number Publication date
CN114945961B (zh) 2024-03-08
CN114945961A (zh) 2022-08-26

Similar Documents

Publication Publication Date Title
US11794748B2 (en) Vehicle system for recognizing objects
US11842282B2 (en) Neural networks for coarse- and fine-object classifications
US20230089380A1 (en) Neural network construction method and apparatus
US10748061B2 (en) Simultaneous localization and mapping with reinforcement learning
JP7263216B2 (ja) ワッサースタイン距離を使用する物体形状回帰
JP7203224B2 (ja) 開放された車両ドアを検出するための分類器のトレーニング
CN109109863B (zh) 智能设备及其控制方法、装置
JP2022506404A (ja) 車両速度を決定する方法及び装置
CN114323054B (zh) 自动驾驶车辆行驶轨迹的确定方法、装置及电子设备
US11610423B2 (en) Spatio-temporal-interactive networks
CN113383283B (zh) 感知信息处理方法、装置、计算机设备和存储介质
WO2023082850A1 (zh) 行人轨迹预测方法、装置及存储介质
WO2022099526A1 (zh) 变道预测回归模型训练方法、变道预测方法和装置
US20230334842A1 (en) Training instance segmentation neural networks through contrastive learning
US20210319287A1 (en) Predicting occupancy probabilities of surrounding agents
CN116858253A (zh) 一种适用室内环境的轻量性预测式导航方法及***
WO2023242223A1 (en) Motion prediction for mobile agents
US20240202393A1 (en) Motion planning
CN113158539A (zh) 交通参与者的长期轨迹预测的方法
CN115943400B (zh) 基于时间与空间学习的轨迹预测方法、装置和计算机设备
KR102465312B1 (ko) 좌표계 기반 이동객체 충돌 예측 시스템 및 방법
RU2801781C1 (ru) Способ и устройство для классификации объекта
Zangirolami et al. Dealing with uncertainty: Balancing exploration and exploitation in deep recurrent reinforcement learning
Sharma et al. MapsTP: HD Map Images Based Multimodal Trajectory Prediction for Automated Vehicles
CN113627562A (zh) 一种目标检测方法、装置、设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20961083

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20961083

Country of ref document: EP

Kind code of ref document: A1