WO2022241771A1 - 一种预测方法、装置、车辆和智能驾驶*** - Google Patents

一种预测方法、装置、车辆和智能驾驶*** Download PDF

Info

Publication number
WO2022241771A1
WO2022241771A1 PCT/CN2021/095235 CN2021095235W WO2022241771A1 WO 2022241771 A1 WO2022241771 A1 WO 2022241771A1 CN 2021095235 W CN2021095235 W CN 2021095235W WO 2022241771 A1 WO2022241771 A1 WO 2022241771A1
Authority
WO
WIPO (PCT)
Prior art keywords
target vehicle
lane
vehicle
target
feature
Prior art date
Application number
PCT/CN2021/095235
Other languages
English (en)
French (fr)
Inventor
李飞
李向旭
范时伟
Original Assignee
华为技术有限公司
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 华为技术有限公司 filed Critical 华为技术有限公司
Priority to PCT/CN2021/095235 priority Critical patent/WO2022241771A1/zh
Publication of WO2022241771A1 publication Critical patent/WO2022241771A1/zh

Links

Images

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W60/00—Drive control systems specially adapted for autonomous road vehicles

Definitions

  • the present invention relates to the technical field of intelligent driving, in particular to a prediction method, device, vehicle and intelligent driving system.
  • the intelligent driving system can be divided into four key functional modules: positioning, environment perception, path planning and decision-making control.
  • the prediction functions such as the road that the target vehicle is about to drive and the trajectory of pedestrians are mainly concentrated in the environment perception module.
  • the high-precision maps used in the prior art belong to structural data.
  • the structured map data is rasterized, and high-dimensional rasterized images are used to represent map information.
  • the network convolutional neural networks, CNN extracts map information features, which makes the calculation complexity higher.
  • the embodiments of the present application provide a prediction method, device, vehicle and intelligent driving system.
  • the present application provides a prediction method, including: obtaining the position information of the own vehicle and the state information of each obstacle within the set range around the own vehicle, each obstacle including the target vehicle; according to the state of each obstacle Information and at least one environmental variable of the target vehicle, extracting the characteristics of each environmental variable, the environmental variable is a variable obtained by processing the first map, the environmental variable includes at least one lane and the semantic element corresponding to the at least one lane Information, the first map is a map determined according to the position information of the own vehicle and the set range on the stored map; according to the characteristics of each environmental variable and the interaction characteristics of the target vehicle, at least one driving direction of the target vehicle is predicted The trajectory, the interaction feature of the target vehicle is the feature of the interaction between the state information of the target vehicle and the state information of other obstacles, and the other obstacles are obstacles other than the target vehicle among the various obstacles.
  • the feasible domain of the future trajectory of the target vehicle can be divided into airspace and There are two layers in the behavior domain, so as to more reasonably describe the uncertainty of the future trajectory of the target vehicle.
  • different environmental variables describe which road and lane in the map the target vehicle’s future trajectory may choose.
  • Multiple trajectories generated under the same environmental variable constraints describe different behaviors of the target, such as acceleration, deceleration, constant speed, and lane change. Multimodal trajectory prediction for target vehicles with better interpretability.
  • the method after obtaining the state information of each obstacle within the set range around the vehicle, it further includes: inputting the state information of each obstacle into a feature extraction model, and extracting the state information of each obstacle at different times The target feature of the target vehicle; the method also includes: weighted and summing the target feature of the target vehicle at the current observation moment and the target feature of the other obstacle at the current observation moment to obtain the distance between the target vehicle and the other obstacle interactive features.
  • feature extraction is performed on the target vehicle and other obstacles so that target features of corresponding objects can be directly used subsequently, thereby reducing the complexity of calculation.
  • the interaction features are obtained by weighting and summing the target features of the target vehicle and other obstacles, so as to reduce the complexity of calculating the interaction features and improve the calculation efficiency. efficiency.
  • extracting the feature of each environmental variable according to the state information of each obstacle and at least one environmental variable of the target vehicle includes: storing all the objects in at least one lane in the at least one environmental variable The position information input feature extraction model of the point column, extract the point column feature of all point columns of each lane in the at least one environmental variable; the point column features of all point columns of this each lane are weighted and summed, and then The semantic element features corresponding to each lane are spliced on the feature dimension to obtain the lane features of each lane, and the semantic element features corresponding to each lane are extracted by feature extraction of the semantic element information corresponding to each lane Obtaining: splicing the lane features of each lane in the at least one environmental variable and the target features of the obstacles at different times to obtain the features of each environmental variable.
  • the road features and semantic element features are extracted for different environmental variables, and then spliced with the target features of the target vehicle to obtain the environmental variables.
  • the features make the constraint conditions strongly correlated with the road and the target vehicle, and predict the trajectory of the target vehicle more accurately in the follow-up.
  • predicting at least one driving trajectory of the target vehicle according to the characteristics of each environmental variable and the interaction characteristics of the target vehicle includes: using the characteristics of each environmental variable, the interaction characteristics of the target vehicle The features and target features of the target vehicle at the current moment are input into the feature extraction model to determine the trajectory features of the target vehicle under different environmental variable constraints.
  • the constraint information of the environmental variables in the airspace is learned, and the specific Predicted trajectories for different behaviors within the road.
  • the method further includes: inputting at least one predicted driving trajectory of the target vehicle, the characteristics of each environmental variable and the interaction features of the target vehicle into the feature extraction model, and determining that the target vehicle passes through each probability of a driving trajectory.
  • the system or the driver of the vehicle can predict the driving trajectory of the target vehicle more intuitively, so as to take subsequent measures such as rushing and avoiding.
  • the embodiment of the present application provides a prediction device, including: a transceiver unit, used to obtain the position information of the own vehicle and the state information of each obstacle within the set range around the own vehicle, and each obstacle includes the target vehicle
  • the processing unit is used to extract the characteristics of each environmental variable according to the state information of each obstacle and at least one environmental variable of the target vehicle, the environmental variable is a variable obtained by processing the first map, and the environmental variable Including at least one lane and semantic element information corresponding to the at least one lane, the first map is a map determined according to the position information of the own vehicle and the set range on the stored map; and according to the characteristics of each environmental variable and the
  • the interaction feature of the target vehicle is to predict at least one driving trajectory of the target vehicle.
  • the interaction feature of the target vehicle is the feature of the interaction between the state information of the target vehicle and the state information of other obstacles. obstacles other than the target vehicle.
  • the processing unit is specifically configured to input the state information of each obstacle into a feature extraction model, and extract target features of each obstacle at different times; the processing unit is also configured to input the The target feature of the target vehicle at the current observation moment and the target feature of the other obstacle at the current observation moment are weighted and summed to obtain the interaction feature between the target vehicle and the other obstacle.
  • the processing unit is specifically configured to input the position information of all point columns in at least one lane in the at least one environmental variable into the feature extraction model, and extract the position information of each lane in the at least one environmental variable.
  • the point column features of all point columns; the point column features of all point columns of each lane are weighted and summed, and then the semantic element features corresponding to each lane are spliced on the feature dimension to obtain the each lane.
  • Lane feature, the semantic element feature corresponding to each lane is obtained by feature extraction of the semantic element information corresponding to each lane; the lane feature of each lane in the at least one environment variable and the respective obstacles in different
  • the target features at each moment are spliced to obtain the features of each environment variable.
  • the processing unit is specifically configured to input the feature of each environmental variable, the interaction feature of the target vehicle and the target feature of the target vehicle at the current moment into the feature extraction model, and determine that the target vehicle is in different Trajectory features under the constraints of environmental variables.
  • the processing unit is further configured to input at least one predicted driving trajectory of the target vehicle, the characteristics of each environmental variable and the interaction features of the target vehicle into the feature extraction model to determine the target vehicle The probability of passing each driving trajectory.
  • an embodiment of the present application provides a vehicle, including at least one processor, and the processor is configured to execute instructions stored in a memory, so that the vehicle executes various possible implementation embodiments of the first aspect.
  • the embodiment of the present application provides an intelligent driving system, which is used to implement various possible implementation embodiments of the first aspect.
  • the embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed in a computer, the computer is made to execute various possible implementation embodiments of the first aspect.
  • an embodiment of the present application provides a computer program product, which is characterized in that the computer program product stores instructions, and when the instructions are executed by a computer, the computer implements various possible implementations of the first aspect. the embodiment.
  • the embodiment of the present application provides a computing device, including a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, each of the steps in the first aspect can be realized. Examples of possible implementations.
  • FIG. 1 is a schematic structural diagram of an intelligent driving system provided by an embodiment of the present application
  • FIG. 2 is a schematic structural diagram of an environment perception module provided by an embodiment of the present application.
  • FIG. 3 is a schematic diagram of a scene where a vehicle is detected on a road provided by an embodiment of the present application
  • FIG. 4 is a schematic diagram of the environment encoder architecture provided by the embodiment of the present application.
  • FIG. 5 is a schematic diagram of dimensionally splicing target features and lane features provided by the embodiment of the present application.
  • FIG. 6 is a flow chart of a prediction method provided in an embodiment of the present application.
  • FIG. 7 is a schematic diagram of a prediction device provided by an embodiment of the present application.
  • the intelligent driving system uses sensors to detect the surrounding environment and its own state, such as navigation and positioning information, road information, obstacle information such as other vehicles and pedestrians, its own pose information and motion state information, etc. After a certain decision-making and planning algorithm , Precisely control the speed and steering of the vehicle, so that it can drive automatically without the driver's supervision. As shown in FIG. 1 , according to the functional requirements of the intelligent driving system 100 , the system 100 can be divided into a positioning module 10 , an environment perception module 20 , a path planning module 30 and a decision-making control module 40 .
  • the positioning module 10 is used to obtain the location of the vehicle through data collected by sensors such as a global positioning system (global positioning system, GPS) unit, an inertial navigation system (inertial navigation system, INS) unit, an odometer, a camera, and a radar in the sensor system. Location and navigation information.
  • sensors such as a global positioning system (global positioning system, GPS) unit, an inertial navigation system (inertial navigation system, INS) unit, an odometer, a camera, and a radar in the sensor system.
  • GPS global positioning system
  • INS inertial navigation system
  • the positioning technology can be divided into three types: absolute positioning, relative positioning and combined positioning according to the positioning method.
  • Absolute positioning refers to the realization through GPS, that is, the absolute position and heading information of the vehicle on the earth is obtained through satellites;
  • relative positioning refers to the acquisition of acceleration and angular acceleration information through INS, odometer and other sensors according to the initial position and posture of the vehicle, and its By integrating the time, the current pose information relative to the initial pose can be obtained;
  • combined positioning refers to the combination of absolute positioning and relative positioning to make up for the shortcomings of a single positioning method.
  • the environment perception module 20 is used for the data collected by sensors such as GPS unit, INS unit, odometer, camera, radar (lidar, millimeter wave radar, ultrasonic radar, etc.) The location and navigation information of the car, the perception of the environment information around the car, the status information of the car and other vehicles around the car, pedestrians and other objects.
  • sensors such as GPS unit, INS unit, odometer, camera, radar (lidar, millimeter wave radar, ultrasonic radar, etc.)
  • radar lidar, millimeter wave radar, ultrasonic radar, etc.
  • the environmental information can include the shape, direction, curvature, slope, lane, traffic sign, signal light of the road, the position, size, direction and speed of other vehicles or pedestrians, etc.; the state information can include forward speed, acceleration, steering angle, Body position and attitude etc.
  • the path planning module 30 is used to obtain the position and navigation information of the own vehicle through the positioning module 10, and the environment information around the own vehicle sensed by the environment perception module 20, the status information of the own vehicle and other vehicles around the vehicle, pedestrians and other objects, Plan a reasonable driving route for the own vehicle, other vehicles, pedestrians and other objects around the vehicle.
  • path planning it can be divided into global path planning and local path planning.
  • Global path planning refers to planning a global path from the current position of the vehicle to the destination when the global map is known; local path planning refers to changing lanes, turning, avoiding obstacles, etc. according to environmental perception information. Plan a safe and smooth driving path in real time.
  • the decision control module 40 includes a decision function and a control function.
  • the decision-making function is used to obtain data according to the positioning module 10, the environment perception module 20 and the path planning module 30, and decide which lane to choose for the own car or other vehicles, whether to change lanes, whether to follow the car, whether to detour, whether to stop, etc. Behavior;
  • the control function is used to execute the decision-making instructions issued by the decision-making function, control the vehicle to achieve the desired speed and steering angle, and control the turn signal, horn, doors and windows and other components.
  • the process of predicting the trajectories of other vehicles around the vehicle and the trajectories of pedestrians it is generally implemented in the environment perception module 20, and can also be implemented in the path planning module 30, or even in an independent It is realized in the prediction module (not shown in the figure), and is specifically determined according to the usage scenario of the prediction result, which is not limited here.
  • the predicting module is arranged between the environment sensing module 20 and the path planning module 30, and is used to receive the environment information around the own vehicle, the state information of the own vehicle and other vehicles around the vehicle, pedestrians and other objects sensed by the environment sensing module 20, Predict the motion trajectories of vehicles and pedestrians, and then input the predicted motion trajectories into the path planning module 30 .
  • the present application will take the environment perception module 20 to predict a vehicle around the own vehicle as a target, and take the driving track of the target vehicle as an example to describe the technical solution of the present application.
  • the environment perception module 20 shown in FIG. 2 can be divided into a map information processor 201 , an object encoder 202 , an environment encoder 203 , an interaction information processor 204 and a multimodal decoder 205 according to its execution functions.
  • the map information processor 201 is generally connected with memory, positioning device, external terminal, cloud server and other equipment to obtain its high-precision map and the current position information of the own vehicle, and then according to the high-precision map and the current position of the own vehicle Information, process part of the map in a certain area around the self-vehicle in the high-precision map, and obtain different environmental variables.
  • Different environmental variables describe which road and which lane in the map the target vehicle's future trajectory may choose.
  • the same environmental variable Multiple trajectories generated under the constraints describe different behaviors of the target, such as acceleration, deceleration, constant speed, lane change, etc.
  • different environmental variables correspond to different road element information in the map, and each environmental variable includes lanes, stop lines, traffic lights and other semantic traffic sign information.
  • the map information processor 201 determines the part of the map to be processed, it divides the lane lines into different categories according to the topological relationship of the lane lines in the map, and the lane lines of different categories belong to different
  • signs 1, 2, 3, 5 and 6 are straight lanes
  • sign 4 is a right-turning lane (relative to the target vehicle)
  • sign 7 is a left-turning lane (relative to the target vehicle).
  • the first plan is to go straight along the lane marked 3.
  • the second plan is to turn left along the marked lane 3 and then turn left to the marked 7 lane. Change from Lane 3 to Lane 1, then turn right to Lane 4.
  • the target vehicle has three environmental variables, including the lane conditions as follows:
  • Environmental variable 2 mark 3 lanes (identify the road section between the target vehicle of the 3 lanes and the intersecting position with the mark 7 lanes) and mark the 7 lanes;
  • the map information processor 201 determines the corresponding lane-level semantic element information according to the relationship between the map semantic element and the lane, such as traffic light information, left-turn lane, right-turn lane and other information, and then uniquely heats the lane-level semantic element information ( one-hot) encoding, and then added to the environment variable.
  • This application obtains the road elements corresponding to different environmental variables according to the topological structure of the map, which can improve the generalization of the model and the robustness to noise, reduce the dimension of the processed data, and effectively improve the operation efficiency and real-time performance. It avoids rasterizing the map in the prior art, and the rasterized map information encoding method has high complexity and poor encoding accuracy.
  • Status information obtained by this application Generally, it includes position information, speed information, movement direction information and so on. Specific to different types of objects, such as vehicle status information It includes the position information of the vehicle, the speed of the vehicle, the direction information of the vehicle, etc., such as the state information of pedestrians Including pedestrian position information, pedestrian speed, pedestrian orientation information and so on.
  • the target encoder 202 obtains the state information of the target vehicle and each obstacle at different moments in history After that, the status information In the input feature extraction model, extract the state information of the target vehicle and each obstacle at different moments in history Corresponding target features
  • the target encoder 202 uses different historical moments of the target vehicle and each obstacle status information for Input to MLP, MLP projects the state information of the target vehicle and each obstacle at each moment to the high-dimensional feature space, obtains the characteristics of the target vehicle and each obstacle in the high-dimensional space at each moment, and then inputs it into the RNN , extract the features of the state sequence of the target vehicle and each obstacle Representation, the hidden state of the RNN, such that each feature output is associated with historical features preceding that feature.
  • the target encoder 202 can use the formula (1) to express the target feature
  • the encoding process specifically:
  • i represents the serial number of the target
  • t represents different times in history
  • MLP() represents data input into MLP for processing
  • RNN() represents data input into RNN for processing
  • MLP is a forward-structured artificial neural network (ANN), which maps a set of input vectors to a set of output vectors.
  • An MLP can be viewed as a directed graph consisting of multiple layers of nodes, each fully connected to the next layer. Except for the input node, each node is a neuron with a non-linear activation function.
  • the MLP is trained using the supervised learning method of the BP backpropagation (BP) algorithm.
  • BP BP backpropagation
  • the MLP in the feature extraction module is a single-layer fully connected network, the hidden unit dimension is 64, and the activation function is sent to Relu through the batch normalization layer (BN).
  • RNN is a special neural network structure, which is proposed based on the viewpoint that "human cognition is based on past experience and memory". Different from deep neural network (DNN) and CNN, it not only considers the input at the previous moment, but also endows the network with a "memory" function for the previous content.
  • DNN deep neural network
  • RNN is called a recurrent neural network
  • the reason why RNN is called a recurrent neural network is that the current output of a sequence is also related to the previous output.
  • the specific manifestation is that the network will remember the previous information and apply it to the calculation of the current output, that is, the nodes between the hidden layers are no longer connected but connected, and the input of the hidden layer not only includes the output of the input layer Also includes the output of the hidden layer at the previous moment.
  • RNN uses a gated recurrent unit (GRU), and the hidden unit dimension is 128.
  • GRU gated recurrent unit
  • the environment encoder 203 performs encoding according to the received environmental variables and target features, and obtains the features of the lane and the interaction features between the target vehicle and various obstacles, so as to predict the trajectory of the target vehicle more accurately. As shown in FIG. 4 , the environment encoder 203 can be divided into a lane encoding module 2031 and an environment variable estimation module 2032 according to the functions performed.
  • the lane encoding module 2031 After acquiring the different environmental variables sent by the map information processor 201, the lane encoding module 2031 obtains each lane included in each environmental variable and all semantic element information related to each lane.
  • the lane on the stored high-precision map is represented by a series of point columns, which can be understood as the center point of the lane. Therefore, after the lane encoding module 2031 obtains different environmental variables, it inputs the point sequence of each lane into MLP and RNN, and obtains the point sequence feature l j, k of the lane by encoding the point sequence of the lane, using the formula (2) Represents the process of encoding all point column features l j, k of each lane, specifically:
  • P j represents the lane point vector matrix of the j-th lane
  • P j,k represents the k-th point column of the j-th lane
  • l j,k represents the point column at the k-th point column of the j-th lane feature.
  • the lane encoding module 2031 inputs all the semantic element information related to the lane carried in each environmental variable into the MLP, and by encoding it, obtains the semantic element feature e j of all the semantic element information related to each lane.
  • e j represents all semantic element encoding feature vectors related to the jth lane.
  • the lane encoding module 2031 After the lane encoding module 2031 obtains the feature lj, k of all the lanes, through the attention mechanism, the feature lj, k of all the point columns in each lane is weighted and summed to obtain the lane j.
  • the weighted summation refers to taking the attention coefficient output by the attention mechanism module as the weight, and summing the features l j, k of all lane points.
  • lane 1 has 5 lane points
  • the obtained lane j and the semantic element features e j of all semantic element information related to the corresponding lane are spliced on the feature dimension C to obtain the lane feature L j of each lane after splicing, which can be used in formula (3)
  • ⁇ k is the attention weight of the k-th point column
  • d is the hidden unit dimension of the MLP query layer.
  • the environmental variable estimation module 2032 considers the interaction between obstacles such as other vehicles and pedestrians and the lane, so by acquiring the target feature sent by the target encoder 202 and the lane feature L j of each lane after splicing sent by the lane encoding module 2031, through the attention mechanism, the target feature Splicing with the lane feature L j to obtain the environmental variable feature E n and the probability distribution d n of the environmental variable feature.
  • splicing refers to the target feature and the lane feature L j are spliced in the feature dimension to become a feature with a larger dimension.
  • the target feature Take the splicing with the lane feature L j in the feature dimension as an example.
  • the environmental variable estimation module 2032 can use formula (4) to represent the process of environmental variable characteristic E n encoding, specifically:
  • E n is the characteristic of the nth environmental variable
  • ⁇ j, n is the influence coefficient of lane j on the nth environmental variable, that is, the probability of lane intention
  • I j is the lane j after considering the interaction information of other vehicles on lane j feature
  • a j is the interaction feature of all obstacles on lane j
  • ⁇ j, i is the attention weight coefficient of the i-th obstacle on the j-th lane.
  • the environmental variable estimation module 2032 can use formula (5) to represent the process of encoding the probability distribution d i of environmental variable characteristics, specifically:
  • d n represents the probability of the nth environmental variable.
  • the interactive information processor 204 receives the target feature extracted by the target encoder 202 Finally, through the attention mechanism, the target features of the target vehicle Target features with individual obstacles Perform weighted summation to obtain the interaction characteristics between the target vehicle and each obstacle
  • the interaction feature can be represented by formula (6)
  • the encoding process specifically:
  • d i is the weight coefficient of the attention mechanism of the i-th obstacle to the self-vehicle, is the interaction feature of other obstacles to the target vehicle at the current moment.
  • the multimodal decoder 205 receives the environmental variable feature E n extracted by the environment encoder 203 and the interaction feature extracted by the interaction information processor 204 and the target features of the target vehicle Finally, through MLP and RNN in turn, the input features are encoded to obtain multiple predicted trajectories on , t, and the process of predicting the predicted trajectories on, t can be expressed by formula (7), specifically:
  • h n, t-1 is the hidden state of the predicted target at the time t-1 under the constraint of the nth environmental variable
  • o n, t-1 is the decoding output of the predicted target at the time t-1 under the constraint of the nth environmental variable , including c predicted positions at time t-1, c is the number of predicted trajectories under the constraints of a single environmental variable
  • E n is the characteristic of the nth environmental variable, is the feature of mutual information, and the target encoding feature Hidden state as the initial moment of RNN.
  • the multi-modal decoder 205 is then based on the predicted multiple predicted trajectories o n, k , environmental variable features E n and interaction information Through MLP, the probability p n , k corresponding to each predicted trajectory o n, k can be estimated, and formula (8) can be used to express the process of encoding the probability p n, k corresponding to each predicted trajectory, specifically:
  • p n, k is the probability of the k-th predicted trajectory of the predicted target under the constraint of the n-th environmental variable at time t
  • o n, k is the k-th predicted trajectory of the predicted target under the constraint of the n-th environmental variable
  • ⁇ ( ⁇ ) is the sigmoid function
  • the feasible domain constructed by the multiple predicted trajectories on,k predicted by the multimodal decoder 205 can be layered into the air domain and the behavioral domain, so as to more reasonably describe the uncertainty of the future trajectory of the target vehicle .
  • the airspace refers to the road that the target vehicle may travel in the future
  • the behavior domain refers to the behavior that the target vehicle may take in a specific road, such as acceleration, deceleration, lane change, etc.
  • FIG. 6 is a flow chart of a prediction method provided by an embodiment of the present application.
  • the prediction method shown in Figure 6, the specific implementation process is as follows:
  • Step S601 acquiring the position information of the own vehicle and the state information of each obstacle within the set range around the own vehicle.
  • the location information refers to the real-time positioning information of the self-car collected by GPS unit, INS unit and other sensors during driving; each obstacle can refer to other vehicles, pedestrians and other objects; The position information (relative to the own vehicle), speed information, movement direction information, etc. of each obstacle collected by sensors such as accelerometers.
  • Step S602 according to the state information of each obstacle and at least one environmental variable of the target vehicle, the feature of each environmental variable is extracted.
  • the environmental variables are obtained by processing part of the map in a certain area around the vehicle in the high-precision map and the current location information of the self-vehicle in the high-precision map, generally including lanes, stop lines, traffic lights and other semantic traffic sign information .
  • each environmental variable when predicting the trajectory of the target vehicle, includes the corresponding lanes in a driving scheme of the target vehicle, and the corresponding lane-level semantic element information of each lane, such as traffic light information, left turn, right Transfer information, etc.
  • the state information of the target vehicle and other obstacles is first input into the feature extraction model to extract the target features of the target vehicle and other obstacles; Input the point columns of each lane into the feature extraction model to extract the features of each point column of each lane, and input the semantic element information corresponding to each lane into the feature extraction model to extract the semantic element features of each lane; then The features of each point column of each lane are weighted and summed through the attention mechanism, and are spliced with the semantic element features of the corresponding lane in the feature dimension to obtain the lane features of each lane; finally, the lane features of each lane, the target vehicle The target features of other obstacles are weighted and summed through the attention mechanism to obtain the features of each environmental variable.
  • Step S603 predicting at least one driving trajectory of the target vehicle according to the characteristics of each environmental variable and the interaction characteristics of the target vehicle.
  • the interaction feature of the target vehicle is obtained by weighting and summing the target feature of the target vehicle and the target features of each obstacle through the attention mechanism.
  • the characteristics of various environmental variables, the interaction characteristics of the target vehicle and the target features of the target vehicle are input into the feature extraction model to obtain multiple predicted trajectories of the target vehicle; then the multiple predicted trajectories of the target vehicle, Each environmental variable feature and the interaction feature of the target vehicle are input into the feature extraction model to obtain the probability of each predicted trajectory.
  • the feasible domain of the target's future trajectory is divided into two layers: the air domain and the behavioral domain.
  • Different environmental variables describe which road and road in the map the target vehicle's future trajectory may choose.
  • Lane multiple trajectories generated under the same environmental variable constraints describe the different behaviors of the target, such as acceleration, deceleration, constant speed, and lane change, so that the subsequent multimodal trajectory prediction of the target vehicle has better interpretability.
  • map information is vectorized, and simple MLP, RNN, and attention mechanisms are used to encode map information, and the interaction between targets and the interaction between targets and maps is modeled to effectively improve prediction accuracy and computing efficiency.
  • technical solution of the present application can also estimate the probability distribution of environmental variables, and the intended road and intended lane of the target vehicle can be obtained according to the environmental variables. This intention can adapt to various road structures and accurately describe the target behavior.
  • FIG. 7 is a schematic diagram of a prediction device provided by an embodiment of the present application.
  • the apparatus 700 includes a transceiver unit 701 and a processing unit 702 .
  • the specific implementation process of the device 700 is as follows:
  • the transceiver unit 701 is used to obtain the position information of the own vehicle and the state information of each obstacle within the set range around the own vehicle, and each obstacle includes the target vehicle;
  • the processing unit 702 is used to At least one environmental variable of the vehicle, extracting the characteristics of each environmental variable, the environmental variable is a variable obtained by processing the first map, the environmental variable includes at least one lane and the semantic element information corresponding to the at least one lane, the first A map is a map determined according to the position information of the own vehicle and the set range on the stored map; and predicting at least one driving trajectory of the target vehicle according to the characteristics of each environmental variable and the interaction characteristics of the target vehicle, the The interaction feature of the target vehicle is a feature of the interaction between the state information of the target vehicle and the state information of other obstacles, and the other obstacles are obstacles other than the target vehicle among the various obstacles.
  • the processing unit 702 is specifically configured to input the state information of each obstacle into a feature extraction model, and extract target features of each obstacle at different times; the processing unit 702 is also configured to The target feature of the target vehicle at the current observation moment and the target feature of the other obstacle at the current observation moment are weighted and summed to obtain the interaction feature between the target vehicle and the other obstacle.
  • the processing unit 702 is specifically configured to input the position information of all point columns in at least one lane in the at least one environmental variable into the feature extraction model, and extract the position information of each lane in the at least one environmental variable.
  • the point column features of all point columns; the point column features of all point columns of each lane are weighted and summed, and then the semantic element features corresponding to each lane are spliced on the feature dimension to obtain the each lane.
  • Lane feature, the semantic element feature corresponding to each lane is obtained by feature extraction of the semantic element information corresponding to each lane; the lane feature of each lane in the at least one environment variable and the respective obstacles in different
  • the target features at each moment are spliced to obtain the features of each environment variable.
  • the processing unit 702 is specifically configured to input the feature of each environmental variable, the interaction feature of the target vehicle and the target feature of the target vehicle at the current moment into the feature extraction model, and determine that the target vehicle is in different Trajectory features under the constraints of environmental variables.
  • the processing unit 702 is further configured to input at least one predicted driving trajectory of the target vehicle, the characteristics of each environmental variable and the interaction features of the target vehicle into the feature extraction model to determine the target vehicle The probability of passing each driving trajectory.
  • the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed in a computer, the computer is instructed to execute any one of the above-mentioned methods.
  • the present invention provides a computing device, including a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, any one of the above-mentioned methods is realized.
  • computer-readable media may include, but are not limited to: magnetic storage devices (e.g., hard disks, floppy disks, or tapes, etc.), optical disks (e.g., compact discs (compact discs, CDs), digital versatile discs (digital versatile discs, DVDs), etc.), smart cards and flash memory devices (for example, erasable programmable read-only memory (EPROM), card, stick or key drive, etc.).
  • magnetic storage devices e.g., hard disks, floppy disks, or tapes, etc.
  • optical disks e.g., compact discs (compact discs, CDs), digital versatile discs (digital versatile discs, DVDs), etc.
  • smart cards and flash memory devices for example, erasable programmable read-only memory (EPROM), card, stick or key drive, etc.
  • various storage media described herein can represent one or more devices and/or other machine-readable media for storing information.
  • the term "machine-readable medium” may include, but is not limited to, wireless channels and various other media capable of storing, containing and/or carrying instructions and/or data.
  • the prediction apparatus 700 in FIG. 7 may be fully or partially implemented by software, hardware, firmware or any combination thereof.
  • software When implemented using software, it may be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions.
  • the computer program instructions When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application will be generated in whole or in part.
  • the computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable device.
  • the computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transferred from a website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.) to another website site, computer, server or data center.
  • the computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrated with one or more available media.
  • the available medium may be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk (SSD)).
  • sequence numbers of the above-mentioned processes do not mean the order of execution, and the order of execution of the processes should be determined by their functions and internal logic, and should not The implementation process of the embodiment of the present application constitutes no limitation.
  • the disclosed systems, devices and methods may be implemented in other ways.
  • the device embodiments described above are only illustrative.
  • the division of the units is only a logical function division. In actual implementation, there may be other division methods.
  • multiple units or components can be combined or can be Integrate into another system, or some features may be ignored, or not implemented.
  • the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.
  • the unit described as a separate component may or may not be physically separated, and the component shown as a unit may or may not be a physical unit, that is, it may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • this function is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.
  • the technical solution of the embodiment of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions to make a computer device (which may be a personal computer, a server, or an access network device, etc.) execute all or part of the steps of the method in each embodiment of the embodiment of the present application.
  • the aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc and other media that can store program codes. .

Landscapes

  • Engineering & Computer Science (AREA)
  • Automation & Control Theory (AREA)
  • Human Computer Interaction (AREA)
  • Transportation (AREA)
  • Mechanical Engineering (AREA)
  • Traffic Control Systems (AREA)

Abstract

本申请提供了一种预测方法、装置、车辆和智能驾驶***，涉及智能驾驶技术领域。其中，该方法通过获取自车的位置信息和周围设定范围内各个障碍物的状态信息，然后根据各个障碍物的状态信息和目标车辆的至少一个环境变量，提取出每个环境变量的特征，最后根据每个环境变量的特征和目标车辆的交互特征，预测目标车辆的至少一条驾驶轨迹。本申请通过引入环境变量，并结合根据目标车辆与其它障碍物之间的交互特征，来预测目标车辆的未来行驶轨迹，可以实现将目标车辆的未来轨迹的可行域划分为空域和行为域两层，从而更加合理的描述目标车辆未来轨迹的不确定性。

Description

一种预测方法、装置、车辆和智能驾驶*** 技术领域
本发明涉及智能驾驶技术领域，尤其涉及一种预测方法、装置、车辆和智能驾驶***。
背景技术
随着智能化的发展和普及，车辆的智能驾驶成为当前比较热门的研究方向。智能驾驶***根据功能需求，可分为定位、环境感知、路径规划和决策控制四个关键功能模块。其中，目标车辆即将行驶的道路、行人运动轨迹等预测功能主要集中在环境感知模块中。
近些年来，随着传感器硬件发展和算法的进步带来的自动驾驶领域快速发展，对车辆轨迹预测的研究也逐渐增多，目前较为常见的采用多模态预测方法来目标车辆的运动轨迹。也即通过引入服从特定分布的随机隐变量，对随机隐变量进行采样并输入至解码器中生成未来轨迹。但是这种随机隐变量与目标行为之间无明显关系，可解释性较差。
另外，现有技术中使用的高精地图属于结构性数据，通常将结构性的地图数据进行栅格化，用高维栅格化图像表示地图信息，利用通过计算复杂度较高的卷积神经网络(convolutional neural networks，CNN)提取地图信息特征，使得计算的复杂度较高。
发明内容
为了解决上述的问题，本申请的实施例提供了一种预测方法、装置、车辆和智能驾驶***。
第一方面，本申请提供一种预测方法，包括：获取自车的位置信息和自车周围设定范围内各个障碍物的状态信息，该各个障碍物包括目标车辆；根据该各个障碍物的状态信息和该目标车辆的至少一个环境变量，提取出每个环境变量的特征，该环境变量为对第一地图进行处理得到的变量，该环境变量包括至少一个车道和该至少一个车道对应的语义元素信息，该第一地图为存储的地图上根据该自车的位置信息和设定范围确定的地图；根据该每个环境变量的特征和该目标车辆的交互特征，预测该目标车辆的至少一条驾驶轨迹，该目标车辆的交互特征为该目标车辆的状态信息与其它障碍物的状态信息之间交互的特征，该其它障碍物为该各个障碍物中除该目标车辆以外的障碍物。
在该实施方式中，通过引入环境变量，并结合根据目标车辆与其它障碍物之间的交互特征，来预测目标车辆的未来行驶轨迹，可以实现将目标车辆的未来轨迹的可行域划分为空域和行为域两层，从而更加合理的描述目标车辆未来轨迹的不确定性。而且不同环境变量描述目标车辆未来轨迹可能选取地图中的哪条道路和车道，同一环境变量约束下生成的多条轨迹描述目标不同的行为，例如加速、减速、匀速以及换道等行为，使得后续对目标车辆的多模态轨迹预测具有更好的可解释性。
在一种实施方式中，该获取车辆周围设定范围内的各个障碍物的状态信息之后，还包括：将该各个障碍物的状态信息输入特征提取模型中，提取该各个障碍物在不同时刻下的目标特征；该方法还包括：将该目标车辆在当前观测时刻下的目标特征与该其它障碍物在 当前观测时刻下的目标特征进行加权求和，得到该目标车辆与该其它障碍物之间的交互特征。
在该实施方式中，通过对目标车辆和其它各个障碍物进行特征提取，以便后续直接使用对应物体的目标特征，降低计算的复杂度。同时，考虑到其它障碍物对目标车辆在未来行驶过程中的影响，通过目标车辆与其它障碍物的目标特征进行加权求和来得到交互特征，实现降低计算交互特征过程中的复杂度和提高计算效率。
在一种实施方式中，该根据该各个障碍物的状态信息和该目标车辆的至少一个环境变量，提取出每个环境变量的特征，包括：将该至少一个环境变量中至少一个车道中的所有点列的位置信息输入特征提取模型中，提取出该至少一个环境变量中每个车道的所有点列的点列特征；将该每个车道的所有点列的点列特征进行加权求和，然后与该每个车道对应的语义元素特征在特征维度上进行拼接，得到该每个车道的车道特征，该每个车道对应的语义元素特征是通过将该每个车道对应的语义元素信息进行特征提取得到；将该至少一个环境变量中该每个车道的车道特征和该各个障碍物在不同时刻下的目标特征进行拼接，得到每个环境变量的特征。
在该实施方式中，为了便于得到对预测轨迹进行约束的环境变量的特征，通过对不同的环境变量进行道路特征和语义元素特征的提取，然后再与目标车辆的目标特征拼接，得到的环境变量特征，使得约束条件与道路和目标车辆之间强相关，在后续预测目标车辆的轨迹更加精准。
在一种实施方式中，该根据该每个环境变量的特征和该目标车辆的交互特征，预测该目标车辆的至少一条驾驶轨迹，包括：将该每个环境变量的特征、该目标车辆的交互特征和该目标车辆当前时刻的目标特征输入特征提取模型中，确定该目标车辆在不同环境变量约束下的轨迹特征。
在该实施方式中，通过将每个环境变量的特征、目标车辆的交互特征和目标车辆当前时刻的目标特征输入特征提取模型中，如RNN，学习环境变量在空域的约束信息，预测出在特定道路内的不同行为的预测轨迹。
在一种实施方式中，该方法还包括：将该目标车辆的至少一条预测的驾驶轨迹、该每个环境变量的特征和该目标车辆的交互特征输入特征提取模型中，确定该目标车辆通过每个驾驶轨迹的概率。
在该实施方式中，通过预测出每个驾驶轨迹的概率，可以让***或自车驾驶员更加直观的预知目标车辆的行驶轨迹，以便采取后续抢行、避让等措施。
第二方面，本申请实施例提供了一种预测装置，包括：收发单元，用于获取自车的位置信息和自车周围设定范围内各个障碍物的状态信息，该各个障碍物包括目标车辆；处理单元，用于根据该各个障碍物的状态信息和该目标车辆的至少一个环境变量，提取出每个环境变量的特征，该环境变量为对第一地图进行处理得到的变量，该环境变量包括至少一个车道和该至少一个车道对应的语义元素信息，该第一地图为存储的地图上根据该自车的位置信息和设定范围确定的地图；以及根据该每个环境变量的特征和该目标车辆的交互特征，预测该目标车辆的至少一条驾驶轨迹，该目标车辆的交互特征为该目标车辆的状态信息与其它障碍物的状态信息之间交互的特征，该其它障碍物为该各个障碍物中除该目标车辆以外的障碍物。
在一种实施方式中，该处理单元，具体用于将该各个障碍物的状态信息输入特征提取模型中，提取该各个障碍物在不同时刻下的目标特征；该处理单元，还用于将该目标车辆在当前观测时刻下的目标特征与该其它障碍物在当前观测时刻下的目标特征进行加权求和，得到该目标车辆与该其它障碍物之间的交互特征。
在一种实施方式中，该处理单元，具体用于将该至少一个环境变量中至少一个车道中的所有点列的位置信息输入特征提取模型中，提取出该至少一个环境变量中每个车道的所有点列的点列特征；将该每个车道的所有点列的点列特征进行加权求和，然后与该每个车道对应的语义元素特征在特征维度上进行拼接，得到该每个车道的车道特征，该每个车道对应的语义元素特征是通过将该每个车道对应的语义元素信息进行特征提取得到；将该至少一个环境变量中该每个车道的车道特征和该各个障碍物在不同时刻下的目标特征进行拼接，得到每个环境变量的特征。
在一种实施方式中，该处理单元，具体用于将该每个环境变量的特征、该目标车辆的交互特征和该目标车辆当前时刻的目标特征输入特征提取模型中，确定该目标车辆在不同环境变量约束下的轨迹特征。
在一种实施方式中，该处理单元，还用于将该目标车辆的至少一条预测的驾驶轨迹、该每个环境变量的特征和该目标车辆的交互特征输入特征提取模型中，确定该目标车辆通过每个驾驶轨迹的概率。
第三方面，本申请实施例提供了一种车辆，包括至少一个处理器，该处理器用于执行存储器中存储的指令，以使得车辆执行如第一方面各个可能实现的实施例。
第四方面，本申请实施例提供了一种智能驾驶***，用于执行如第一方面各个可能实现的实施例。
第五方面，本申请实施例提供了一种计算机可读存储介质，其上存储有计算机程序，当该计算机程序在计算机中执行时，令计算机执行如第一方面各个可能实现的实施例。
第六方面，本申请实施例提供了一种计算机程序产品，其特征在于，所述计算机程序产品存储有指令，所述指令在由计算机执行时，使得所述计算机实施如第一方面各个可能实现的实施例。
第七方面，本申请实施例提供了一种计算设备，包括存储器和处理器，其特征在于，该存储器中存储有可执行代码，该处理器执行该可执行代码时，实现如第一方面各个可能实现的实施例。
附图说明
下面对实施例或现有技术描述中所需使用的附图作简单地介绍。
图1为本申请实施例提供的一种智能驾驶***的架构示意图；
图2为本申请实施例提供的环境感知模块的架构示意图；
图3为本申请实施例提供的检测车辆所处道路的场景示意图；
图4为本申请实施例提供的环境编码器架构示意图；
图5为本申请实施例提供的目标特征和车道特征在维度上进行拼接示意图；
图6为本申请实施例提供的一种预测方法的流程图；
图7为本申请实施例提供的一种预测装置的示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行描述。
智能驾驶***，就是利用传感器检测周围环境和自身状态，如导航定位信息、道路信息、其他车辆和行人等障碍物信息、自身的位姿信息及运动状态信息等等，经过一定的决策规划算法后，精确的控制车辆行驶速度和转向，从而现实在不需要驾驶员的监控下即自动驾驶。如图1所示，根据智能驾驶***100的功能需求，可将该***100分为定位模块10、环境感知模块20、路径规划模块30和决策控制模块40。
定位模块10用于通过传感器***中的全球定位***(global positioning system，GPS)单元、惯性导航***(inertial navigation system，INS)单元、里程计、摄像头、雷达等传感器采集的数据，获取自车的位置和导航信息。
其中，定位技术按照定位的方式可分为绝对定位、相对定位和组合定位三种。绝对定位是指通过GPS实现，即通过卫星获得车辆在地球上的绝对位置和航向信息；相对定位是指根据车辆的初始位姿，通过INS、里程计等传感器获得加速度和角加速度信息，将其对时间进行积分，即可得到相对初始位姿的当前位姿信息；组合定位是指将绝对定位和相对定位结合，以弥补单一定位方式的不足。
环境感知模块20用于通过传感器***中的GPS单元、INS单元、里程计、摄像头、雷达(激光雷达、毫米波雷达、超声波雷达等等)等传感器采集的数据，以及定位模块10获取的自车的位置和导航信息，感知自车周围的环境信息、自车和自车周围其它车辆、行人等物体的状态信息。
其中，环境信息可以包括道路的形状、方向、曲率、坡度、车道，交通标志，信号灯，其他车辆或行人的位置、大小、前进方向和速度等；状态信息可以包括前进速度、加速度、转向角度、车身位置及姿态等。
路径规划模块30用于通过定位模块10获取的自车的位置和导航信息，以及环境感知模块20感知的自车周围的环境信息、自车和自车周围其它车辆、行人等物体的状态信息，为自车、和自车周围其它车辆、行人等物体规划合理的行驶路线。其中，根据路径规划的范围，可分为全局路径规划和局部路径规划。全局路径规划是指在已知全局地图的情况下，从车辆当前位置规划出一条到目的地的全局路径；局部路径规划指根据环境感知信息，在换道、转弯、躲避障碍物等情况下，实时规划出一条安全、平顺的行驶路径。
决策控制模块40包括决策功能和控制功能。其中，决策功能用于根据定位模块10、环境感知模块20和路径规划模块30得到数据，决定自车或其它车辆选取哪条车道、是否换道、是否跟车行驶、是否绕行、是否停车等行为；控制功能用于执行决策功能下发的决策指令，控制车辆达到期望的速度和转向角度，以及对转向灯、喇叭、门窗等部件的控制。
本申请实施例中，对自车周围的其它车辆的行驶轨迹和行人的运动轨迹进行预测过程中，一般在环境感知模块20中实现，也可以在路径规划模块30中实现，甚至在一个独立的预测模块(图中未示出)中实现，具体根据预测结果使用场景来确定，在此不做限定。其中，预测模块设置在环境感知模块20与路径规划模块30之间，用于接收环境感知模块20感知自车周围的环境信息、自车和自车周围其它车辆、行人等物体的状态信息后，对车辆和行人的运动轨迹进行预测，然后将预测的运动轨迹输入到路径规划模块30。
本申请下面将以环境感知模块20预测自车周围的一辆车作为目标，对该目标车辆的行驶轨迹为例来讲述本申请的技术方案。
如图2所示的环境感知模块20，其根据执行功能，可划分为地图信息处理器201、目标编码器202、环境编码器203、交互信息处理器204和多模态解码器205。
地图信息处理器201一般与存储器、定位装置、外部的终端、云服务器等设备连接，用于获取其高精度地图和自车当前时刻的位置信息，然后根据高精度地图和自车当前时刻的定位信息，对高精度地图中自车周围一定区域内的部分地图进行处理，得到不同的环境变量，不同的环境变量描述目标车辆未来轨迹可能选取地图中的哪条道路和哪个车道，同一个环境变量约束下生成的多条轨迹描述目标的不同的行为，如加速、减速、匀速、换道等行为。其中，不同的环境变量对应地图中不同道路元素信息，每个环境变量包括车道、停车线、红绿灯以及其它语意交通标志信息。
示例性地，如图3所示，地图信息处理器201确定将要处理的部分地图后，根据该地图中车道线的拓扑关系，将车道线划分为不同的类别，不同类别的车道线属于不同的道路，标识1、标识2、标识3、标识5和标识6为直行车道，标识4为右转弯车道(相对于目标车辆)，标识7为左转弯车道(相对于目标车辆)。对于目标车辆来说，其未来有三种行驶方案，第一方案，沿着标识3的车道直行，第二方案，沿着标识3车道后向左转到标识7车道上，第三方案，从标识3车道转换到标识1车道上，然后向右转到标识4车道上。
所以目标车辆有三个环境变量，分别包括车道情况为：
环境变量1：标识1车道、标识2车道和标识3车道；
环境变量2：标识3车道(标识3车道目标车辆和与标识7车道相交位置之间的路段)和标识7车道；
环境变量3：标识1车道(目标车辆的位置和切换到标识1车道的最短缓冲位置之间的路段)、标识1车道(目标车辆标识1车道与标识4车道相交位置之间的路段)和标识4车道。
另外，地图信息处理器201再根据地图语义元素与车道关系，确定相应的车道级语意元素信息，如红绿灯信息、左转道、右转道等信息，然后将车道级语意元素信息进行独热(one-hot)编码，再添加到环境变量中。
本申请根据地图的拓扑结构得到不同环境变量对应的道路元素，可以提高模型的泛化性以及对噪声的鲁棒性，降低处理数据的维度，有效提高运算效率与实时性。避免了现有技术中将地图进行栅格化，栅格化后的地图信息编码方式的复杂度高且编码精度差。
目标编码器202通过传感器***、定位模块10和环境感知模块20等模块获取自车周围所有车辆(包括目标车辆)、行人等障碍物在历史上不同时刻的状态信息
Figure PCTCN2021095235-appb-000001
其中，i＝0表示目标车辆，i＞0表示除目标车辆以外的其它车辆、行人等障碍物，t表示历史上不同时刻。
本申请获取的状态信息
Figure PCTCN2021095235-appb-000002
一般包括位置信息、速度信息、运动方向信息等等。具体到不同类型的物体，如车辆的状态信息
Figure PCTCN2021095235-appb-000003
则包括车辆的位置信息、车辆的速度、车辆行驶的方向信息等等，如行人的状态信息
Figure PCTCN2021095235-appb-000004
包括行人的位置信息、行人的速度、行人的朝向信息等等。
目标编码器202在得到目标车辆和各个障碍物的历史上不同时刻的状态信息
Figure PCTCN2021095235-appb-000005
后，将状态信息
Figure PCTCN2021095235-appb-000006
输入特征提取模型中，提取目标车辆和各个障碍物的历史上不同时刻的状态信息
Figure PCTCN2021095235-appb-000007
对应的目标特征
Figure PCTCN2021095235-appb-000008
示例性地，以特征提取模型为多层感知机(multi layer perceptron，MLP)和循环神经网络(recurrent neural network，RNN)为例，目标编码器202将目标车辆和各个障碍物的历史上不同时刻的状态信息
Figure PCTCN2021095235-appb-000009
输入到MLP，MLP对目标车辆和各个障碍物的每一时刻的状态信息投影至高维特征空间，得到在高维度空间中的目标车辆和各个障碍物的每一时刻的特征，再输入到RNN中，提取出目标车辆和各个障碍物的状态序列的特征
Figure PCTCN2021095235-appb-000010
表示，即RNN的隐状态，使得输出的每个特征均与该特征之前的历史特征相关联。其中，目标编码器202可以用公式(1)表示目标特征
Figure PCTCN2021095235-appb-000011
编码的过程，具体为：
Figure PCTCN2021095235-appb-000012
其中，i表示目标的序号，i＝0表示目标车辆，t表示历史上不同时刻，MLP()表示数据输入MLP中进行处理，RNN()表示数据输入RNN中进行处理，
Figure PCTCN2021095235-appb-000013
表示目标车辆的特征。
MLP是一种前向结构的人工神经网络(artificial neural network，ANN)，映射一组输入向量到一组输出向量。MLP可以被看做是一个有向图，由多个节点层组成，每一层全连接到下一层。除了输入节点，每个节点都是一个带有非线性激活函数的神经元。使用BP反向传播(back propagation，BP)算法的监督学习方法来训练MLP。本申请中，特征提取模块中的MLP为单层全连接网络，隐单元维度为64，经过批归一化层(BN)送入激活函数为Relu。
RNN是一种特殊的神经网络结构，它是根据“人的认知是基于过往的经验和记忆”这一观点提出的。与深度神经网络(deep neural network，DNN)和CNN不同的是，它不仅考虑前一时刻的输入，而且赋予了网络对前面的内容的一种“记忆”功能。RNN之所以称为循环神经网路，即一个序列当前的输出与前面的输出也有关。具体的表现形式为网络会对前面的信息进行记忆并应用于当前输出的计算中，即隐藏层之间的节点不再无连接而是有连接的，并且隐藏层的输入不仅包括输入层的输出还包括上一时刻隐藏层的输出。本申请中，RNN采用是门控循环单元(gated recurrent unit，GRU)，隐单元维度为128。
环境编码器203根据接收到的环境变量和目标特征进行编码，得到车道的特征和目标车辆与各个障碍物之间的交互特征，以便后续对目标车辆的运动轨迹预测更加精准。如图4所示，环境编码器203根据执行的功能，可分为车道编码模块2031和环境变量估计模块2032。
车道编码模块2031获取到地图信息处理器201发送的不同环境变量后，得到每个环境变量中包括的各个车道和每个车道相关的所有语意元素信息。
通常情况下，存储的高精度地图上的车道是由一系列点列来表示，可以理解为车道中心点。所以车道编码模块2031得到不同环境变量后，将每个车道的点列输入到MLP和RNN中，通过对车道的点列进行编码，得到车道的点列特征l j，k，用公式(2)表示每个车道的所有的点列特征l j，k编码的过程，具体为：
Figure PCTCN2021095235-appb-000014
其中，P j表示第j条车道的车道点向量矩阵，P j，k表示第j条车道的第k个点列，l j，k表示第j条车道的第k个点列处的点列特征。
同时，车道编码模块2031将每个环境变量中携带的车道相关的所有语意元素信息输入到MLP中，通过对其进行编码，得到每个车道相关的所有语意元素信息的语意元素特征e j。其中，e j表示第j条车道相关的所有语意元素编码特征向量。
车道编码模块2031在得到所有车道的点列特征l j，k后，通过注意力(attention)机制，将每一条车道内所有点列的特征l j，k进行加权求和，得到车道j。其中，加权求和是指将attention机制模块输出的attention系数作为权重，对所有车道点的特征l j，k求和。示例性地，车道1有5个车道点，attention机制输出5个车道点对应的attention系数为α k，，且k＝1，2，3，4，5，加权求和的结果为
Figure PCTCN2021095235-appb-000015
然后，将得到的车道j与对应车道的相关的所有语意元素信息的语意元素特征e j在特征维度C上进行拼接，得到拼接后的每一条车道的车道特征L j，可以用公式(3)表示车道特征L j编码的过程，具体为：
Figure PCTCN2021095235-appb-000016
其中，α k为第k个点列的attention权重，d为MLP query层的隐单元维度。
环境变量估计模块2032考虑到其它车辆、行人等障碍物与车道之间交互，所以通过获取目标编码器202发送的目标特征
Figure PCTCN2021095235-appb-000017
和车道编码模块2031发送的拼接后的每一条车道的车道特征L j，通过attention机制，将目标特征
Figure PCTCN2021095235-appb-000018
和车道特征L j进行拼接，得到环境变量特征E n和环境变量特征的概率分布d n。
其中，拼接是指将目标特征
Figure PCTCN2021095235-appb-000019
和车道特征L j在特征维度上拼接成为维度更大的特征。示例性地，如图5所示，以目标特征
Figure PCTCN2021095235-appb-000020
和车道特征L j在特征维度上进行拼接为例。其中，目标特征
Figure PCTCN2021095235-appb-000021
在特征维度上为(W 1，H 1，C 1)，W 1＝1，H 1＝1；车道特征L j在特征维度上为(W 2，H 2，C 2)，且W 2＝1，H 2＝1。在拼接时，可以直接将目标特征
Figure PCTCN2021095235-appb-000022
和车道特征L j在特征维度C上进行拼接，得到拼接后的特征为(W，H，C 1+C 2)，且W＝1，H＝1。
环境变量估计模块2032可以用公式(4)表示环境变量特征E n编码的过程，具体为：
Figure PCTCN2021095235-appb-000023
其中，E n为第n个环境变量的特征；γ j，n为车道j在第n个环境变量的影响系数，即车道意图概率；I j为考虑其他车辆对车道j交互信息后的车道j特征；A j为所有障碍物对车道j的交互特征；β j，i为第i个障碍物对第j个车道的注意力权重系数。
环境变量估计模块2032可以用公式(5)表示环境变量特征的概率分布d i编码的过程，具体为：
Figure PCTCN2021095235-appb-000024
其中，d n表示第n个环境变量的概率。
交互信息处理器204通过接收目标编码器202提取的目标特征
Figure PCTCN2021095235-appb-000025
后，通过attention机制，将目标车辆的目标特征
Figure PCTCN2021095235-appb-000026
与各个障碍物的目标特征
Figure PCTCN2021095235-appb-000027
进行加权求和，得到目标车辆与各个障碍物之间的交互特征
Figure PCTCN2021095235-appb-000028
可以用公式(6)表示交互特征
Figure PCTCN2021095235-appb-000029
编码的过程，具体为：
Figure PCTCN2021095235-appb-000030
其中，d i为其第i个障碍物对自车的attention机制的权重系数，
Figure PCTCN2021095235-appb-000031
为其它障碍物对目标车辆当前时刻的交互特征。
多模态解码器205通过接收环境编码器203提取的环境变量特征E n、交互信息处理器204提取的交互特征
Figure PCTCN2021095235-appb-000032
和目标车辆的目标特征
Figure PCTCN2021095235-appb-000033
后，依次通过MLP和RNN中，对输入的特征进行编码，得到多条预测轨迹o n，t，可以用公式(7)表示预测轨迹o n，t编码预测的过程，具体为：
Figure PCTCN2021095235-appb-000034
其中，h n，t-1为预测目标在第n个环境变量约束下t-1时刻的隐态，o n，t-1为预测目标第n个环境变量约束下t-1时刻的解码输出，包含t-1时刻的c个预测位置，c为单个环境变量约束下的预测轨迹个数；E n为第n个环境变量特征，
Figure PCTCN2021095235-appb-000035
为交互信息的特征，目标编码特征
Figure PCTCN2021095235-appb-000036
作为RNN的初始时刻的隐态。
多模态解码器205再根据预测的多条预测轨迹o n，k、环境变量特征E n和交互信息
Figure PCTCN2021095235-appb-000037
通过MLP，估计每条预测轨迹o n，k对应的概率p n，k，可以用公式(8)表示每条预测轨迹对应的概率p n，k编码的过程，具体为：
Figure PCTCN2021095235-appb-000038
其中，p n，k为在t时刻下，预测目标在第n个环境变量约束下第k条预测轨迹的概率，o n，k为预测目标在第n个环境变量约束下第k条预测轨迹，σ(·)是sigmoid函数。
本申请实施例中，多模态解码器205预测出的多条预测轨迹o n，k构建的可行域，可分层为空域和行为域，从而更加合理的描述目标车辆未来轨迹的不确定性。其中，空域是指目 标车辆未来可能行驶的道路，行为域是指目标车辆在特定道路内可能采取的行为，例如加速、减速、换道等等。
图6为本申请实施例提供的一种预测方法的流程图。如图6所示的预测方法，具体实现过程如下：
步骤S601，获取自车的位置信息和自车周围设定范围内各个障碍物的状态信息。
其中，位置信息是指自车在行驶过程中通过GPS单元、INS单元等传感器实时采集自车的定位信息；各个障碍物可以指其它车辆、行人等物体；状态信息是指通过相机、激光雷达、加速度计等传感器采集到的各个障碍物的位置信息(相对于自车)、速度信息、运动方向信息等等。
步骤S602，根据各个障碍物的状态信息和目标车辆的至少一个环境变量，提取出每个环境变量的特征。
其中，环境变量是根据高精度地图和自车当前时刻的定位信息，对高精度地图中车辆周围一定区域内的部分地图进行处理得到的，一般包括车道、停车线、红绿灯以及其它语意交通标志信息。本申请中，对目标车辆进行轨迹预测时，每个环境变量包括目标车辆的一种行驶方案中对应的各个车道，以及各个车道相应的车道级语意元素信息，如红绿灯信息、左转道、右转道等信息。
在提取每个环境变量过程中，先将目标车辆和其它障碍物的状态信息输入到特征提取模型中，以提取出目标车辆和其它障碍物的目标特征；然后将每个环境变量中的各个车道的点列输入到特征提取模型中，以提取出各个车道的各个点列的特征，并将各个车道对应的语意元素信息输入到特征提取模型中，以提取出各个车道的语意元素特征；接着将各个车道的各个点列的特征通过attention机制进行加权求和，并与对应车道的语意元素特征在特征维度上拼接，得到每个车道的车道特征；最后，将每个车道的车道特征、目标车辆和其它障碍物的目标特征通过attention机制进行加权求和得到每个环境变量特征。
步骤S603，根据每个环境变量的特征和目标车辆的交互特征，预测目标车辆的至少一条驾驶轨迹。其中，目标车辆的交互特征是将目标车辆的目标特征与各个障碍物的目标特征通过attention机制进行加权求和得到的。
在预测轨迹过程中，将各个环境变量特征、目标车辆的交互特征和目标车辆的目标特征输入到特征提取模型中，得到目标车辆的多条预测轨迹；然后再将目标车辆的多条预测轨迹、各个环境变量特征和目标车辆的交互特征输入到特征提取模型中，得到每条预测轨迹的概率。
本申请实施例中，通过引入环境变量与多模态解码器，将目标未来轨迹的可行域划分为空域和行为域两层，不同环境变量描述目标车辆未来轨迹可能选取地图中的哪条道路和车道，同一环境变量约束下生成的多条轨迹描述目标不同的行为，例如加速、减速、匀速以及换道等行为，使得后续对目标车辆的多模态轨迹预测具有更好的可解释性。
另外，将地图信息矢量化，采用简单的MLP、RNN和attention机制对地图信息编码，对目标间的交互以及目标与地图间的交互建模，有效提高预测精度与运算效率。而且本申请技术方案还能够估计环境变量的概率分布，根据环境变量可以得到目标车辆的意图道路与意图车道，这种意图能够适应各种道路结构，可以对目标行为进行准确描述。
图7为本申请实施例提供的一种预测装置的示意图。如图7所示，该装置700包括收发单元701和处理单元702。该装置700具体实现过程如下：
收发单元701用于获取自车的位置信息和自车周围设定范围内各个障碍物的状态信息，该各个障碍物包括目标车辆；处理单元702用于根据该各个障碍物的状态信息和该目标车辆的至少一个环境变量，提取出每个环境变量的特征，该环境变量为对第一地图进行处理得到的变量，该环境变量包括至少一个车道和该至少一个车道对应的语义元素信息，该第一地图为存储的地图上根据该自车的位置信息和设定范围确定的地图；以及根据该每个环境变量的特征和该目标车辆的交互特征，预测该目标车辆的至少一条驾驶轨迹，该目标车辆的交互特征为该目标车辆的状态信息与其它障碍物的状态信息之间交互的特征，该其它障碍物为该各个障碍物中除该目标车辆以外的障碍物。
在一种实施方式中，该处理单元702具体用于将该各个障碍物的状态信息输入特征提取模型中，提取该各个障碍物在不同时刻下的目标特征；该处理单元702还用于将该目标车辆在当前观测时刻下的目标特征与该其它障碍物在当前观测时刻下的目标特征进行加权求和，得到该目标车辆与该其它障碍物之间的交互特征。
在一种实施方式中，该处理单元702具体用于将该至少一个环境变量中至少一个车道中的所有点列的位置信息输入特征提取模型中，提取出该至少一个环境变量中每个车道的所有点列的点列特征；将该每个车道的所有点列的点列特征进行加权求和，然后与该每个车道对应的语义元素特征在特征维度上进行拼接，得到该每个车道的车道特征，该每个车道对应的语义元素特征是通过将该每个车道对应的语义元素信息进行特征提取得到；将该至少一个环境变量中该每个车道的车道特征和该各个障碍物在不同时刻下的目标特征进行拼接，得到每个环境变量的特征。
在一种实施方式中，该处理单元702具体用于将该每个环境变量的特征、该目标车辆的交互特征和该目标车辆当前时刻的目标特征输入特征提取模型中，确定该目标车辆在不同环境变量约束下的轨迹特征。
在一种实施方式中，该处理单元702还用于将该目标车辆的至少一条预测的驾驶轨迹、该每个环境变量的特征和该目标车辆的交互特征输入特征提取模型中，确定该目标车辆通过每个驾驶轨迹的概率。
本发明提供一种计算机可读存储介质，其上存储有计算机程序，当该计算机程序在计算机中执行时，令计算机执行上述任一项方法。
本发明提供一种计算设备，包括存储器和处理器，该存储器中存储有可执行代码，该处理器执行该可执行代码时，实现上述任一项方法。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请实施例的范围。
此外，本申请实施例的各个方面或特征可以实现成方法、装置或使用标准编程和/或工 程技术的制品。本申请中使用的术语“制品”涵盖可从任何计算机可读器件、载体或介质访问的计算机程序。例如，计算机可读介质可以包括，但不限于:磁存储器件(例如，硬盘、软盘或磁带等)，光盘(例如，压缩盘(compact disc，CD)、数字通用盘(digital versatile disc，DVD)等)，智能卡和闪存器件(例如，可擦写可编程只读存储器(erasable programmable read-only memory，EPROM)、卡、棒或钥匙驱动器等)。另外，本文描述的各种存储介质可代表用于存储信息的一个或多个设备和/或其它机器可读介质。术语“机器可读介质”可包括但不限于，无线信道和能够存储、包含和/或承载指令和/或数据的各种其它介质。
在上述实施例中，图7中的预测装置700可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。该计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行该计算机程序指令时，全部或部分地产生按照本申请实施例该的流程或功能。该计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。该计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，该计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。该计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。该可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘Solid State Disk(SSD))等。
应当理解的是，在本申请实施例的各种实施例中，上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的***、装置和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的***、装置和方法，可以通过其它的方式实现。例如，以上所描述的装置实施例仅仅是示意性的，例如，该单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个***，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
该作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
该功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实施例的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者接入网设备等)执行本申请实施例各个实施例该方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、 随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
以上该，仅为本申请实施例的具体实施方式，但本申请实施例的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请实施例揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请实施例的保护范围之内。

Claims (15)

  1. 一种预测方法，其特征在于，包括：
    获取自车的位置信息和自车设定范围内各个障碍物的状态信息，所述各个障碍物包括目标车辆；
    根据所述各个障碍物的状态信息和所述目标车辆的至少一个环境变量，提取出每个环境变量的特征，所述环境变量为对第一地图进行处理得到的变量，所述环境变量包括至少一个车道和所述至少一个车道对应的语义元素信息，所述第一地图为存储的地图上根据所述自车的位置信息和设定范围确定的地图；
    根据所述每个环境变量的特征和所述目标车辆的交互特征，预测所述目标车辆的至少一条驾驶轨迹，所述目标车辆的交互特征为所述目标车辆的状态信息与其它障碍物的状态信息之间交互的特征，所述其它障碍物为所述各个障碍物中除所述目标车辆以外的障碍物。
  2. 根据权利要求1所述的方法，其特征在于，所述获取车辆周围设定范围内的各个障碍物的状态信息之后，还包括：
    将所述各个障碍物的状态信息输入特征提取模型中，提取所述各个障碍物在不同时刻下的目标特征；
    所述方法还包括：
    将所述目标车辆在当前观测时刻下的目标特征与所述其它障碍物在当前观测时刻下的目标特征进行加权求和，得到所述目标车辆与所述其它障碍物之间的交互特征。
  3. 根据权利要求2所述的方法，其特征在于，所述根据所述各个障碍物的状态信息和所述目标车辆的至少一个环境变量，提取出每个环境变量的特征，包括：
    将所述至少一个环境变量中至少一个车道中的所有点列的位置信息输入特征提取模型中，提取出所述至少一个环境变量中每个车道的所有点列的点列特征；
    将所述每个车道的所有点列的点列特征进行加权求和，然后与所述每个车道对应的语义元素特征在特征维度上进行拼接，得到所述每个车道的车道特征，所述每个车道对应的语义元素特征是通过将所述每个车道对应的语义元素信息进行特征提取得到；
    将所述至少一个环境变量中所述每个车道的车道特征和所述各个障碍物在不同时刻下的目标特征进行拼接，得到每个环境变量的特征。
  4. 根据权利要求2或3所述的方法，其特征在于，所述根据所述每个环境变量的特征和所述目标车辆的交互特征，预测所述目标车辆的至少一条驾驶轨迹，包括：
    将所述每个环境变量的特征、所述目标车辆的交互特征和所述目标车辆当前时刻的目标特征输入特征提取模型中，确定所述目标车辆在不同环境变量约束下的轨迹特征。
  5. 根据权利要求1-4任意一项所述的方法，其特征在于，所述方法还包括：
    将所述目标车辆的至少一条预测的驾驶轨迹、所述每个环境变量的特征和所述目标车辆的交互特征输入特征提取模型中，确定所述目标车辆通过每个驾驶轨迹的概率。
  6. 一种预测装置，其特征在于，包括：
    收发单元，用于获取自车的位置信息和自车周围设定范围内各个障碍物的状态信息，所述各个障碍物包括目标车辆；
    处理单元，用于根据所述各个障碍物的状态信息和所述目标车辆的至少一个环境变量， 提取出每个环境变量的特征，所述环境变量为对第一地图进行处理得到的变量，所述环境变量包括至少一个车道和所述至少一个车道对应的语义元素信息，所述第一地图为存储的地图上根据所述自车的位置信息和设定范围确定的地图；以及
    根据所述每个环境变量的特征和所述目标车辆的交互特征，预测所述目标车辆的至少一条驾驶轨迹，所述目标车辆的交互特征为所述目标车辆的状态信息与其它障碍物的状态信息之间交互的特征，所述其它障碍物为所述各个障碍物中除所述目标车辆以外的障碍物。
  7. 根据权利要求6所述的装置，其特征在于，所述处理单元，具体用于
    将所述各个障碍物的状态信息输入特征提取模型中，提取所述各个障碍物在不同时刻下的目标特征；
    所述处理单元，还用于
    将所述目标车辆在当前观测时刻下的目标特征与所述其它障碍物在当前观测时刻下的目标特征进行加权求和，得到所述目标车辆与所述其它障碍物之间的交互特征。
  8. 根据权利要求7所述的装置，其特征在于，所述处理单元，具体用于
    将所述至少一个环境变量中至少一个车道中的所有点列的位置信息输入特征提取模型中，提取出所述至少一个环境变量中每个车道的所有点列的点列特征；
    将所述每个车道的所有点列的点列特征进行加权求和，然后与所述每个车道对应的语义元素特征在特征维度上进行拼接，得到所述每个车道的车道特征，所述每个车道对应的语义元素特征是通过将所述每个车道对应的语义元素信息进行特征提取得到；
    将所述至少一个环境变量中所述每个车道的车道特征和所述各个障碍物在不同时刻下的目标特征进行拼接，得到每个环境变量的特征。
  9. 根据权利要求7或8所述的装置，其特征在于，所述处理单元，具体用于
    将所述每个环境变量的特征、所述目标车辆的交互特征和所述目标车辆当前时刻的目标特征输入特征提取模型中，确定所述目标车辆在不同环境变量约束下的轨迹特征。
  10. 根据权利要求6-9任意一项所述的装置，其特征在于，所述处理单元，还用于
    将所述目标车辆的至少一条预测的驾驶轨迹、所述每个环境变量的特征和所述目标车辆的交互特征输入特征提取模型中，确定所述目标车辆通过每个驾驶轨迹的概率。
  11. 一种车辆，包括至少一个处理器，所述处理器用于执行存储器中存储的指令，以使得车辆执行如权利要求1-5任一所述的方法。
  12. 一种智能驾驶***，用于执行如权利要求1-5中的任一项所述的方法。
  13. 一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行权利要求1-5中任一项的所述的方法。
  14. 一种计算机程序产品，其特征在于，所述计算机程序产品存储有指令，所述指令在由计算机执行时，使得所述计算机实施权利要求1-5任意一项所述的方法。
  15. 一种计算设备，包括存储器和处理器，其特征在于，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现权利要求1-5中任一项所述的方法。
PCT/CN2021/095235 2021-05-21 2021-05-21 一种预测方法、装置、车辆和智能驾驶*** WO2022241771A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/095235 WO2022241771A1 (zh) 2021-05-21 2021-05-21 一种预测方法、装置、车辆和智能驾驶***

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/095235 WO2022241771A1 (zh) 2021-05-21 2021-05-21 一种预测方法、装置、车辆和智能驾驶***

Publications (1)

Publication Number Publication Date
WO2022241771A1 true WO2022241771A1 (zh) 2022-11-24

Family

ID=84141087

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/095235 WO2022241771A1 (zh) 2021-05-21 2021-05-21 一种预测方法、装置、车辆和智能驾驶***

Country Status (1)

Country Link
WO (1) WO2022241771A1 (zh)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190072973A1 (en) * 2017-09-07 2019-03-07 TuSimple Data-driven prediction-based system and method for trajectory planning of autonomous vehicles
CN110989636A (zh) * 2020-02-26 2020-04-10 北京三快在线科技有限公司 一种障碍物的轨迹预测方法及装置
CN111319615A (zh) * 2020-03-16 2020-06-23 湖北亿咖通科技有限公司 智能代客泊车方法、计算机可读存储介质及电子设备
CN111750886A (zh) * 2020-05-12 2020-10-09 北京智行者科技有限公司 局部路径规划方法及装置
CN111873995A (zh) * 2020-08-04 2020-11-03 禾多科技（北京）有限公司 高速公路自动驾驶上下匝道的***及方法
CN112068545A (zh) * 2020-07-23 2020-12-11 哈尔滨工业大学（深圳） 一种无人驾驶车辆在十字路口的行驶轨迹规划方法、***及存储介质
CN112417756A (zh) * 2020-11-13 2021-02-26 清华大学苏州汽车研究院（吴江） 一种自动驾驶算法的交互式仿真测试***

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190072973A1 (en) * 2017-09-07 2019-03-07 TuSimple Data-driven prediction-based system and method for trajectory planning of autonomous vehicles
CN110989636A (zh) * 2020-02-26 2020-04-10 北京三快在线科技有限公司 一种障碍物的轨迹预测方法及装置
CN111319615A (zh) * 2020-03-16 2020-06-23 湖北亿咖通科技有限公司 智能代客泊车方法、计算机可读存储介质及电子设备
CN111750886A (zh) * 2020-05-12 2020-10-09 北京智行者科技有限公司 局部路径规划方法及装置
CN112068545A (zh) * 2020-07-23 2020-12-11 哈尔滨工业大学（深圳） 一种无人驾驶车辆在十字路口的行驶轨迹规划方法、***及存储介质
CN111873995A (zh) * 2020-08-04 2020-11-03 禾多科技（北京）有限公司 高速公路自动驾驶上下匝道的***及方法
CN112417756A (zh) * 2020-11-13 2021-02-26 清华大学苏州汽车研究院（吴江） 一种自动驾驶算法的交互式仿真测试***

Similar Documents

Publication Publication Date Title
US11131993B2 (en) Methods and systems for trajectory forecasting with recurrent neural networks using inertial behavioral rollout
US20210362596A1 (en) End-To-End Tracking of Objects
CN111971574B (zh) 用于自动驾驶车辆的lidar定位的基于深度学习的特征提取
US11531346B2 (en) Goal-directed occupancy prediction for autonomous driving
CN111771135B (zh) 自动驾驶车辆中使用rnn和lstm进行时间平滑的lidar定位
US10997729B2 (en) Real time object behavior prediction
JP2022016419A (ja) 軌跡予測方法及び装置
US11827214B2 (en) Machine-learning based system for path and/or motion planning and method of training the same
CN111771141A (zh) 自动驾驶车辆中使用3d cnn网络进行解决方案推断的lidar定位
JP2022549913A (ja) 知覚システム
CN114556375A (zh) 跨批次归一化
CN116348938A (zh) 用于预测动态对象行为的方法和***
JP2023505426A (ja) センサデータアラインメントの補正及び環境マッピング関連アプリケーション
US20240241257A1 (en) Out of sequence preception system
Wang et al. Deep understanding of big geospatial data for self-driving: Data, technologies, and systems
Masmoudi et al. Autonomous car-following approach based on real-time video frames processing
WO2024049925A1 (en) Trajectory prediction based on a decision tree
US20240101150A1 (en) Conditional trajectory determination by a machine learned model
WO2022241771A1 (zh) 一种预测方法、装置、车辆和智能驾驶***
US20220326714A1 (en) Unmapped u-turn behavior prediction using machine learning
US11292462B1 (en) Object trajectory from wheel direction
US20240208546A1 (en) Predictive models for autonomous vehicles based on object interactions
RU2808469C2 (ru) Способ управления роботизированным транспортным средством
US20240217548A1 (en) Trajectory prediction for autonomous vehicles using attention mechanism
US20240208536A1 (en) Cascaded trajectory refinement

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: 21940224

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: 21940224

Country of ref document: EP

Kind code of ref document: A1