WO2017210901A1 - 车辆自动驾驶的速度规划方法、装置及计算装置 - Google Patents

车辆自动驾驶的速度规划方法、装置及计算装置 Download PDF

Info

Publication number
WO2017210901A1
WO2017210901A1 PCT/CN2016/085310 CN2016085310W WO2017210901A1 WO 2017210901 A1 WO2017210901 A1 WO 2017210901A1 CN 2016085310 W CN2016085310 W CN 2016085310W WO 2017210901 A1 WO2017210901 A1 WO 2017210901A1
Authority
WO
WIPO (PCT)
Prior art keywords
partition
decision
result
speed
machine learning
Prior art date
Application number
PCT/CN2016/085310
Other languages
English (en)
French (fr)
Inventor
周小成
姜岩
彭进展
周鑫
张丹
罗赛
Original Assignee
驭势科技（北京）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 驭势科技（北京）有限公司 filed Critical 驭势科技（北京）有限公司
Priority to PCT/CN2016/085310 priority Critical patent/WO2017210901A1/zh
Priority to CN201680001425.1A priority patent/CN107182206B/zh
Priority to ES16904354T priority patent/ES2833674T3/es
Priority to EP16904354.4A priority patent/EP3460613B1/en
Priority to US16/308,353 priority patent/US10564644B2/en
Publication of WO2017210901A1 publication Critical patent/WO2017210901A1/zh
Priority to US16/735,960 priority patent/US11747818B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
    • G05D1/00—Control of position, course, altitude or attitude of land, water, air or space vehicles, e.g. using automatic pilots
    • G05D1/02—Control of position or course in two dimensions
    • G05D1/021—Control of position or course in two dimensions specially adapted to land vehicles
    • G05D1/0212—Control of position or course in two dimensions specially adapted to land vehicles with means for defining a desired trajectory
    • G05D1/0223—Control of position or course in two dimensions specially adapted to land vehicles with means for defining a desired trajectory involving speed control of the vehicle
    • 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
    • B60W60/001—Planning or execution of driving tasks
    • 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
    • B60W30/00—Purposes of road vehicle drive control systems not related to the control of a particular sub-unit, e.g. of systems using conjoint control of vehicle sub-units
    • 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
    • B60W30/00—Purposes of road vehicle drive control systems not related to the control of a particular sub-unit, e.g. of systems using conjoint control of vehicle sub-units
    • B60W30/14—Adaptive cruise control
    • B60W30/143—Speed control
    • B60W30/146—Speed limiting
    • 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
    • B60W30/00—Purposes of road vehicle drive control systems not related to the control of a particular sub-unit, e.g. of systems using conjoint control of vehicle sub-units
    • B60W30/14—Adaptive cruise control
    • B60W30/16—Control of distance between vehicles, e.g. keeping a distance to preceding vehicle
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
    • G05D1/00—Control of position, course, altitude or attitude of land, water, air or space vehicles, e.g. using automatic pilots
    • G05D1/02—Control of position or course in two dimensions
    • G05D1/021—Control of position or course in two dimensions specially adapted to land vehicles
    • G05D1/0212—Control of position or course in two dimensions specially adapted to land vehicles with means for defining a desired trajectory
    • G05D1/0221—Control of position or course in two dimensions specially adapted to land vehicles with means for defining a desired trajectory involving a learning process
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/22—Indexing; Data structures therefor; Storage structures
    • G06F16/2282—Tablespace storage structures; Management thereof
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • 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
    • B60W50/00—Details of control systems for road vehicle drive control not related to the control of a particular sub-unit, e.g. process diagnostic or vehicle driver interfaces
    • B60W2050/0062—Adapting control system settings
    • B60W2050/0075—Automatic parameter input, automatic initialising or calibrating means
    • B60W2050/0083—Setting, resetting, calibration
    • B60W2050/0088—Adaptive recalibration
    • 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
    • B60W2520/00—Input parameters relating to overall vehicle dynamics
    • B60W2520/10—Longitudinal speed
    • 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
    • B60W2554/00—Input parameters relating to objects
    • B60W2554/80—Spatial relation or speed relative to objects
    • B60W2554/801—Lateral distance
    • 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
    • B60W2554/00—Input parameters relating to objects
    • B60W2554/80—Spatial relation or speed relative to objects
    • B60W2554/802—Longitudinal distance
    • 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
    • B60W2554/00—Input parameters relating to objects
    • B60W2554/80—Spatial relation or speed relative to objects
    • B60W2554/804—Relative longitudinal speed
    • 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
    • B60W2555/00—Input parameters relating to exterior conditions, not covered by groups B60W2552/00, B60W2554/00
    • B60W2555/20—Ambient conditions, e.g. wind or rain
    • 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
    • B60W2555/00—Input parameters relating to exterior conditions, not covered by groups B60W2552/00, B60W2554/00
    • B60W2555/60—Traffic rules, e.g. speed limits or right of way
    • 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
    • B60W2720/00—Output or target parameters relating to overall vehicle dynamics
    • B60W2720/10—Longitudinal speed

Definitions

  • the present invention relates to the field of vehicle control, and in particular, to a speed planning method, apparatus, and computing device for automatic driving of a vehicle.
  • Speed planning and control is an important research content in autonomous driving.
  • the basic goal is to plan the expected speed of a series of subsequent time points based on the detected state (such as the current speed, the speed of the preceding vehicle, and the distance from the preceding vehicle). And calculate the final control parameters of the car (such as throttle and brake) to actually control the car.
  • Speed planning ensures the passenger's basic comfort and safety while ensuring that passengers are absolutely safe when they have unintended behaviors such as sudden braking.
  • the effects of the models learned through machine learning methods are directly related to the training data. Since the driver usually drives in a very comfortable range, the collected training data is difficult to cover all possible scenes (such as extreme speeds and extreme distances such as short distances), although generalization, etc. Technology can solve this problem to a certain extent, but it cannot completely solve this problem. In addition, the collected driver's behavior may not fully meet the needs of comfort and safety, such as some drivers are not It can guarantee the distance from the front car, so when the current car suddenly brakes, it is difficult to drive the vehicle to ensure complete braking without hitting the front car. Obviously, models trained with this data can't handle this as well.
  • Another problem with machine learning is that it is difficult to make local fine-tuning of the models it trains.
  • speed planning and control applied to autonomous driving it is usually necessary to locally adjust the model for a particular situation.
  • the adjustment of each parameter learned may bring about the overall uncontrollable influence.
  • the model can also be modified by adding more training cases, but this method has a longer period and the last trained model is not fully predictable.
  • the object of the present invention is to overcome the shortcomings and deficiencies in the prior art, and to provide a new speed planning method, device and computing device for vehicle automatic driving.
  • a speed planning method for automatic driving of a vehicle which specifically includes:
  • each training sample is described by a multi-dimensional feature component forming an input space and a decision result forming an output space, each dimension of the multi-dimensional feature component is a speed planning related variable for describing a state of a particular moment of the vehicle, the decision result indicating an expected speed at a next moment and/or a control parameter value associated with the speed control;
  • real-time decision step real-time obtaining the feature components of the vehicle in motion as the input feature quantity, determining the input partition to which the input feature quantity belongs, and querying the partition decision table based on the determined partition to obtain the corresponding decision result.
  • the speed planning method may further include a real-time control step of issuing a control command to the vehicle based on the obtained decision result, thereby controlling the speed of the vehicle.
  • the partitioning decision result of the partition may be adjusted.
  • partitioning decision result of the partition is adjusted, and the partitioning decision result of the partition may be adjusted according to experience, or the partition learning may be performed by a machine learning method, and the partitioning decision result of the partition may be adjusted.
  • each dimension feature component may include a current vehicle speed, a distance from the preceding vehicle, and a front vehicle relative Speed and maximum speed.
  • the feature space needs to be discretized before the machine learning step.
  • the discretization coding method is preferably Tiling Coding, in which each feature component is encoded with Tiling Coding having only one Tiling, and the dimension in which each feature component is located is divided into A preferred 7-13 interval is used to partition the input space.
  • the spatial size of the discretization coding result obtained by using the discretization coding method needs to be calculated.
  • the dynamic storage method is used to store the partition decision table, and only the traversal training is performed.
  • the input of the space stores the output result of the corresponding decision model, and stores the trained decision model backup in addition to the partition decision table;
  • the static storage method is used to store the partition decision table, and traverse all the codes Space, storing the output of the decision model.
  • the input feature quantity is first discretized and encoded by the discretization coding method, and then the obtained discretization coding result is used as an index of the partition decision table, and when the partition decision table is static When the storage method is stored, directly obtain the stored decision result in the partition decision table;
  • the input feature quantity is first discretized and encoded by the discretization coding method, and then the obtained discretization coding result is used as an index of the partition decision table, and when the partition decision table is dynamic
  • the storage method is stored, if the decision result of the discretization coding result is stored in the partition decision table, the decision result is directly obtained from the partition decision table; if the decision result of the discretization coding result is not stored in the partition decision table, The stored decision model is invoked to obtain a decision result, and the obtained decision result is added to the partition decision table.
  • the method of machine learning may adopt a supervised learning method or an unsupervised learning method, or an enhanced learning method.
  • the two inputs are considered to belong to the same partition.
  • the discretization coding method may include various Coarse Coding methods, such as Tile Coding.
  • the speed planning device may further comprise a real time control unit.
  • the machine learning unit is configured to perform machine learning using the training sample set to obtain a machine learning model, each training sample being described by a multidimensional feature component forming an input space and a decision result forming an output space, each dimension of the multidimensional feature component It is a variable related to speed planning for describing a state of a specific moment of the vehicle, the decision result indicating an expected speed of the next moment and/or a control parameter value related to the speed control.
  • the partition decision table obtaining unit is configured to partition the input space, and based on the obtained machine learning model, obtain a decision result corresponding to the determined partition, and form a partition decision table corresponding to the corresponding decision result.
  • the real-time decision-making unit is configured to obtain each dimension feature component of the running vehicle as an input feature quantity in real time, determine an input partition to which the input feature quantity belongs, and query a partition decision table to obtain a corresponding decision result based on the determined partition.
  • the real-time control unit is configured to issue a control command to the vehicle based on the obtained decision result, thereby controlling the speed of the vehicle.
  • the real-time control unit is further configured to adjust the partitioning decision result of the partition in the real-time control when it is found that the decision result of the determined partition does not meet the expected result.
  • the partition decision table obtaining unit is further configured to adjust the partition decision content of the partition, and the partition decision content of the partition may be adjusted according to experience, or the partition may be learned by a machine learning method, and the The partitioning decision result of the partition.
  • the real-time decision unit is further configured to each dimension feature component including a current vehicle speed, a distance of the preceding vehicle, a front vehicle relative speed, and a maximum vehicle speed.
  • the speed planning device for vehicle automatic driving of the present invention further includes a discretization coding unit configured to discretize encoding the input samples of the training samples and the real-time decision stage.
  • the discretization coding method adopted by the discretization coding unit is Tiling Coding, in which each feature component is encoded with Tiling Coding having only one Tiling, and each feature component is used.
  • the dimensions are divided into 7-13 intervals to partition the input space.
  • the partition decision table obtaining unit is configured to calculate a spatial size of the discretization coding result obtained by using the discretization coding method, and when the space is greater than the determination threshold, the dynamic storage side is adopted.
  • the method stores the partition decision table, only traverses the input of the training space, stores the output result of the corresponding decision model, and also stores the trained decision model for use; when the space is smaller than the determined threshold, the static storage method is used to store the partition decision table. , iterate through all the coding spaces and store the output of the decision model.
  • the real-time decision unit is configured to discretize the input feature quantity by using the discretization coding method, and use the obtained discretization coding result as an index of the partition decision table, and when the partition decision table is stored by using a static storage method When you get the decision result stored in the partition decision table directly.
  • the real-time decision unit is configured to discretize the input feature quantity by using the discretization coding method, and use the obtained discretization coding result as an index of the partition decision table, and when the partition decision table is stored by using a dynamic storage method If the decision result of the discretization coding result is stored in the partition decision table, the decision result is directly obtained from the partition decision table; if the decision result of the discretization coding result is not stored in the partition decision table, the call is performed.
  • the stored decision model is used to obtain the decision results and the resulting decision results are added to the partition decision table.
  • the method of machine learning may be a supervised learning method or an unsupervised learning method, or an enhanced learning method.
  • the partition decision table obtaining unit is configured such that the two inputs belong to the same partition as long as the final discrete codes of the two inputs are the same.
  • the discretization coding method is selected from the group consisting of various Coarse Coding methods, such as Tile Coding.
  • the feedback unit is configured to determine the number of partitions that have been adjusted by the partitioning decision result, and when the number of the partitions exceeds a predetermined threshold, cause the machine learning unit and the partition decision table obtaining unit to perform the machine learning operation and the partition decision again. Table get operation.
  • a computing device for speed planning of automatic driving of a vehicle comprising a storage component and a processor, wherein the storage component stores a set of computer executable instructions when the set of computer executable instructions is When the processor executes, performing the following steps: a machine learning step, using a training sample set for machine learning, obtaining a machine learning model, each training sample being described by a multi-dimensional feature component forming an input space and a decision result forming an output space
  • Each dimension of the multi-dimensional feature component is a variable related to speed planning for describing a state of a specific moment of the vehicle, the decision result indicating an expected speed of the next moment and/or a control parameter value related to the speed control;
  • Decision table acquisition steps partitioning the input space, and based on the obtained machine
  • the learning model obtains the decision result corresponding to the determined partition, forms a partition decision table corresponding to the corresponding decision result of each partition; real-time decision step, real-time obtains each dimension feature component of the moving vehicle as the input
  • the speed planning method, device and computing device provided by the partition decision table technology are suitable for the vehicle automatic driving technology, and the problem that the machine learning training model cannot be locally adjusted is well solved, and a certain partition can be easily modified.
  • the decision is made without affecting the decision results of other partitions at all, thus completing the local adjustment.
  • the intuitive nature of the partition decision table can help to identify and solve problems in the machine learning process.
  • the partition decision table can speed up the decision process, and the partition decision table can get faster decision speed.
  • FIG. 1 is a general flow chart of a speed planning method for automatic driving of a vehicle according to an embodiment of the present invention
  • Figure 2 is a schematic illustration of an operational process and input and output of machine learning training and applications
  • FIG. 3 is a schematic diagram showing an operation process and input and output of machine learning training and application for performing discretization coding processing before applying a machine learning method
  • Figure 4 shows the Tile Coding discretization coding method in the Coarse Coding category.
  • Figure 5 illustrates the process of establishing and optimizing an entire decision partition table in the case of employing discretization coding, in accordance with one embodiment of the present invention.
  • FIG. 6 shows a flow chart of a method of implementing a real-time decision step in accordance with one embodiment of the present invention.
  • FIG. 7 shows a flow diagram of a 700 of a speed planning control method including real-time control steps in accordance with an embodiment of the present invention.
  • FIG. 8 shows a block diagram of a structure of a speed planning device 800 for automatic driving of a vehicle according to an embodiment of the present invention.
  • the inventors have proposed the present invention to separately process the training and application of the machine learning model: during training, we perform training according to a general machine learning method; before applying these trained models, the input space is divided into Multiple partitions are stored and the decision results in the model are stored in different partitions to form a partition-organized decision table. Thus, the decision-making process of applying the model is transformed into a process of checking the partition decision table.
  • the intuitive nature of the partition decision table can help to identify and solve problems in the machine learning process.
  • the partition decision table can speed up the decision process, and the partition decision table can get faster decision speed. For autopilot speed planning with high real-time requirements, it is important to easily adjust local strategies and have quick decision speeds.
  • FIG. 1 shows a general flow chart of a speed planning method for automatic driving of a vehicle according to an embodiment of the present invention.
  • step S110 a machine learning step is performed, machine learning is performed using a training sample set, and a machine learning model is obtained, each training sample being described by a multidimensional feature component forming an input space and a decision result forming an output space, the multidimensional feature component
  • Each dimension of the dimension is a variable relating to speed planning for describing a state of a particular moment of the vehicle, the outcome of the decision indicating an expected speed at the next moment and/or a value of the control parameter associated with the speed control.
  • step S120 a partition decision table obtaining step is performed to partition the input space, and based on the obtained machine learning model, a decision result corresponding to the determined partition is obtained, and a partition decision table in which each partition corresponds to the corresponding decision result is formed.
  • step S130 real-time decision is made, real-time obtaining each dimension feature component of the running vehicle as an input feature quantity, determining an input partition to which the input feature quantity belongs, and querying the partition decision table to obtain a corresponding decision based on the determined partition. result.
  • the training and application of the machine learning model are separately processed in accordance with an embodiment of the present invention.
  • the specific machine learning method may be supervised learning that requires labeling data, unsupervised learning without labeling data, or enhanced learning.
  • the input space is divided into multiple partitions according to the discretization coding method during training, and the decision results in the model are stored in different partitions to form a partition.
  • Organizational decision table. Therefore, the speed planning method for vehicle automatic driving according to the embodiment of the present invention is to convert the decision process of the application model into a process of checking the partition decision table.
  • FIG. 2 schematically illustrates an operational schematic of machine learning training and applications and an exemplary schematic of inputs and outputs.
  • a decision model 230 is obtained, and for the input space partition 240, the partitioning and decision model are combined to obtain a partition decision table 250.
  • partition decision table 250 When making a decision using the partition decision table, for the incoming input 270, it is first partitioned 240. After determining the partition to which it belongs, the partition decision table 250 is queried to obtain the corresponding decision result.
  • partitioning for partitions of continuous dimensions, for example, partitioning can be based on empirical knowledge of vehicle control experts.
  • the input space prior to applying the machine learning method, is discretized to extract features and then processed.
  • the discretization coding method naturally divides the continuous input space into multiple partitions.
  • the partition here is not limited to a partition of the same dimension data contiguous space. As long as the final code is the same, it can be considered that the two samples belong to the same partition; in other words, the partition here is not necessarily continuously divided according to the original space, but Divided according to the results of discrete coding.
  • 3 shows an exemplary schematic diagram of the operation process and input and output of machine learning training and application for performing discretization coding processing before applying the machine learning method, wherein any input, whether it is a training example or an application example, The discretization coding process 380 is performed first.
  • FIG. 4 shows a Tile Coding discretization coding method in a Coarse Coding class.
  • Tiling at three different positions divides the entire space into different small areas, and these small areas can be used as partitions.
  • the input space may be composed of four dimensions of the current vehicle speed, the distance of the preceding vehicle, the relative speed of the preceding vehicle, and the maximum vehicle speed.
  • the front car here is a broad concept and is not limited to vehicles. When there is no object in front, you can virtualize a front car and set the distance and relative speed of the virtual front car. The maximum speed is limited by various conditions (such as road restrictions, weather) The maximum travel speed under the line limit). Therefore, the coding space for speed planning and control in the field of automatic driving is usually small, so it is more suitable for partitioning.
  • each input dimension is encoded with Tiling Coding with only one Tiling, and each input dimension is divided into 10 left and right partitions to meet the speed decision requirements of autonomous driving. For example, for the maximum speed, we can partition it by the partition size of 10 km/h, and divide it into [0,10,20,30,40,50,60,70,80,90,100 according to the actual situation. , 110, 120] A total of 12 partitions. The number of partitions at this time is about 10 4 .
  • partition tables of this size are very easy to store and process. In some extreme cases (such as a lot of input dimensions or a large amount of discretization coding output space), the resulting code space may exceed the storage space of the computer.
  • a hash table or other method of saving storage space can be used to dynamically store and process the partition decision table.
  • step S120 shown in FIG. 1 An example of the establishment and optimization process of the entire decision partition table in the case of employing discretization coding according to an embodiment of the present invention, which may be used to perform step S120 shown in FIG. 1, will be described below with reference to FIG.
  • the threshold in FIG. 5 can be set according to the storage space of the machine. It should be noted that this is only an example. It is not necessary to discretize the training samples before the machine learning training.
  • step S510 the computing machine learning method uses the spatial size of the discretization coding result.
  • step S520 it is determined whether the size of the space is greater than a certain threshold.
  • step S520 When the result of the determination in step S520 is negative, proceeding to step S530, the static storage partition decision table: traverses all the coding spaces, and stores the output result of the decision model; otherwise, when the determination result is YES, the process proceeds to step S540, using the hash. Or other dynamic storage method storage partition decision table: only the input of the training space is traversed, and the output result of the corresponding decision model is stored; then proceeding to step S550, in addition to the partition decision table, the trained decision model is stored.
  • the input is encoded using a discrete encoding method.
  • the input here may be the current vehicle speed obtained in real time during the automatic driving of the vehicle, the distance to the preceding vehicle, the relative speed of the preceding vehicle, and the maximum vehicle speed.
  • the input dimensions and specific feature quantities may vary depending on the speed planning method.
  • step S620 the obtained discretization result is used as an index of the partition block, and the partition decision table is queried.
  • the decision result can be directly obtained from the partition decision table;
  • the partition decision table is stored by the dynamic storage method and the decision result of the discretization coding result is not stored in the partition decision table, the decision result cannot be obtained directly from the partition decision table.
  • step S630 it is determined whether a decision result is obtained. If the answer is "Yes”, then proceeding to step S650, the decision result is returned. On the other hand, if the answer is "NO”, the process proceeds to step S640, the stored decision model is called to obtain a decision result, and the obtained decision result is added to the partition decision table, and then proceeds to step S650.
  • FIG. 6 is an example of a decision process in the case where discrete coding is performed and the case where static storage or dynamic storage is used as before.
  • the process of making a decision using the partition decision table may be different.
  • the partition to which the partition belongs may be directly determined according to the input, and then the query is performed.
  • the partition decision table obtains the decision result; in the case where the decision result is not obtained by querying the partition decision table, in the case where the decision model is stored, the stored decision model may be called to obtain the decision result, or the feedback may be directly indicated.
  • the decision result and so on cannot be given.
  • the partitioning decision result of the partition can be quickly adjusted. There is no need to modify the training use case set or adjust the training parameters to retrain, nor to adjust the parameters of the trained model, but only need to modify the decision result of the corresponding partition on the partition decision table, so as to ensure that the adjustment is limited to this partition. It does not affect the decision results on other partitions.
  • the specific method for adjusting the result of the partition decision may be to adjust the partition decision result of the partition according to experience, or learn the partition by a machine learning method. For example, when Now the current speed is 10 km/h, the distance from the preceding car is 200 m, the relative speed of the preceding car is 5 km/h, and the maximum speed is 30 km/h. When the speed control is abnormal, you can refer to an experienced one.
  • the driver's reaction in this case converts it into a corresponding decision result and stores it in the decision table corresponding to the interval to which the input belongs. Or you can collect data from many drivers while driving, collect all the decision results corresponding to the input space, and get a desired decision result through induction or simple machine learning.
  • FIG. 7 shows a flowchart of a speed planning control method 700 including a real-time control step according to an embodiment of the present invention. Steps S710-S730 in FIG. 7 are similar to steps S110-S130 shown in FIG. 1, and the difference is that The real-time control step S740, that is, based on the obtained decision result, issues a control command to the vehicle to control the speed of the vehicle.
  • real-time control is performed, which is not exclusive control, but can be combined with other control strategies of automatic driving (such as steering control, etc.) to comprehensively control the vehicle.
  • the number of partitions that have been adjusted by the partition decision result may be determined after adjusting the partition decision result of the partition, when the partition is When the number exceeds the predetermined threshold, the overall training step is re-executed, the machine learning is re-executed, and then a new decision model is obtained, and the partition is again performed, that is, step S110 to step S130 shown in FIG. 1 can be re-executed based on the feedback results, that is, The machine learning step and the partition decision table take steps.
  • the machine learning is re-executed, it is possible to try to obtain better machine learning by modifying the training parameters and adding corresponding training cases based on the feedback of the application decision result.
  • a speed planning device for automatic driving of a vehicle is also provided.
  • the speed planning apparatus includes the following units: a machine learning unit 810, a partition decision table obtaining unit 820, and a real-time decision unit 830.
  • a real-time control unit 830 may also be included.
  • the machine learning unit 810 is configured to perform machine learning using a training sample set to obtain a machine learning model, each training sample being described by a multidimensional feature component forming an input space and a decision result forming an output space, each dimension of the multidimensional feature component being A speed planning-related variable for describing a state of a particular time of the vehicle, the decision result indicating an expected speed and/or speed control at the next moment The relevant control parameter values.
  • the partition decision table obtaining unit 820 is configured to partition the input space, and based on the obtained machine learning model, obtain a decision result corresponding to the determined partition, and form a partition decision table in which each partition corresponds to the corresponding decision result.
  • the real-time decision unit 830 is configured to obtain each dimension feature component of the running vehicle as an input feature quantity in real time, determine an input partition to which the input feature quantity belongs, and query the partition decision table to obtain a corresponding decision result based on the determined partition.
  • the real-time control unit 840 is configured to issue a control command to the vehicle based on the obtained decision result, thereby controlling the speed of the vehicle.
  • the speed planning apparatus may further include a local partitioning decision adjustment unit configured to adjust the partitioning decision result of the partition when it is found that the decision result of the determined partition is not in conformity with the expectation.
  • the partition decision content of the partition may be adjusted according to experience, or the partition may be learned by a machine learning method, and the partition decision result of the partition may be adjusted.
  • the speed planning apparatus further includes a discretization coding unit configured to discretize encoding the input samples of the training samples and the real-time decision stage.
  • the input has only four dimensions of the current speed, the distance to the preceding vehicle, the relative speed of the preceding vehicle, and the maximum vehicle speed.
  • the discretization coding method preferably used by the discretization coding unit is Tiling Coding. Only one Tiling Tiling Coding encodes each input dimension, and each input dimension is divided into preferred 7-13 intervals, which is good. Meet the speed decision making needs of autonomous driving.
  • the partition decision table obtaining unit 820 is also configured to calculate the discretization coding method.
  • the resulting spatial size of the discretized coding result may exceed the storage space of the computer.
  • the partition decision table obtaining unit 820 calculates that the space of the discretization coding result obtained by the discretization coding method is larger than the determined threshold, it decides to select the dynamic storage method to store the partition decision table, and at this time, only the input of the training space is traversed. The output of the corresponding decision model is stored, and the trained decision model is stored.
  • the real-time decision unit 830 of the present invention is configured to discretize the input feature quantity by using the discretization coding method, and then use the obtained discretization coding result as a partition.
  • the index of the policy table because the partition decision table is stored by the dynamic storage method at this time, if the decision result of the discretization coding result is stored in the partition decision table, the decision result is directly obtained from the partition decision table; otherwise, if the partition decision table If there is no decision result of storing the discretization coding result, the stored decision model is called to obtain a decision result, and the obtained decision result is added to the partition decision table.
  • the partition decision table obtaining unit 820 of the present invention calculates that the space of the discretization coding result obtained by the discretization coding method is smaller than the determination threshold
  • the static storage method is used to store the partition decision table, traverse all the coding spaces, and store the decision model. Output the result.
  • the real-time decision unit 830 of the present invention is configured to first discretize the input feature quantity by using the discretization coding method, and then use the obtained discretization coding result as an index of the partition decision table, because the partition at this time
  • the decision table is stored by using the static storage method, and the stored decision result in the partition decision table can be directly obtained.
  • the machine learning method that the machine learning unit 810 is configured to employ may be a supervised learning method or an unsupervised learning method, or may be an enhanced learning method.
  • the partition decision table obtaining unit 820 is configured such that the two inputs belong to the same partition as long as the final discrete codes of the two inputs are the same.
  • the discretization coding method configured as the discretization coding unit may be selected from various Coarse Coding methods, such as Tile Coding.
  • the speed planning apparatus 800 may further include a feedback unit configured to determine that the partitioned decision result is adjusted after adjusting the partitioning decision result of the partition when the decision result of the certain partition is found to be unsatisfactory.
  • the number of partitions when the number of the partitions exceeds a predetermined threshold, causes the machine learning unit 810 and the partition decision table obtaining unit 820 to perform the machine learning operation and the partition decision table obtaining operation again.
  • Another embodiment of the present invention also provides a computing device for speed planning of automatic driving of a vehicle, comprising a storage component and a processor, wherein the storage component stores a set of computer executable instructions when the set of computer executable instructions is When the processor executes, performing the following steps: a machine learning step, using a training sample set for machine learning, obtaining a machine learning model, each training sample being described by a multi-dimensional feature component forming an input space and a decision result forming an output space
  • Each dimension of the multi-dimensional feature component is a variable related to speed planning for describing a state of a specific moment of the vehicle, the decision result indicating an expected speed of the next moment and/or a control parameter value related to the speed control;
  • Decision table acquisition steps partitioning the input space, and based on the obtained machine
  • the learning model obtains the decision result corresponding to the determined partition, forms a partition decision table corresponding to the corresponding decision result of each partition; real-time decision step, real-time obtains each dimension feature component of the moving vehicle as
  • steps of the method may be performed locally in the vehicle, in the cloud, or in combination with the local and the cloud, and the storage of the data may also be stored locally, or stored in the cloud, or locally and The cloud is combined for storage.

Landscapes

  • Engineering & Computer Science (AREA)
  • Automation & Control Theory (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Transportation (AREA)
  • Mechanical Engineering (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Remote Sensing (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Human Computer Interaction (AREA)
  • Control Of Driving Devices And Active Controlling Of Vehicle (AREA)
  • Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Game Theory and Decision Science (AREA)
  • Feedback Control In General (AREA)
  • Traffic Control Systems (AREA)

Abstract

用于车辆自动驾驶的速度规划方法、装置及计算装置。方法包括：采用训练样本集进行机器学习，获得机器学习模型（S110）；对于输入空间进行分区，以及基于所述获得的机器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表（S120）；实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果（S130）。很好地解决了不能对机器学习训练出的模型做局部调整的问题，易于修改某个分区的决策，而完全不影响其它分区的决策结果；分区决策表的直观特性可以很好的帮助发现和解决机器学习过程中存在的问题；分区决策表可以加速决策过程。

Description

车辆自动驾驶的速度规划方法、装置及计算装置 技术领域
本发明涉及车辆控制领域，特别是涉及一种用于车辆自动驾驶的速度规划方法、装置及计算装置。
背景技术
随着车辆技术的发展，车辆自动驾驶成为热点研究领域。速度规划和控制是自动驾驶中的一个重要研究内容，其基本目标是根据检测到的状态(如当前车速、前车的车速、与前车的距离)规划出一系列后续时间点车的预期车速，并计算出车的最终控制参数(如油门和刹车)对车进行实际的控制操作。速度规划在保证乘客的基本舒适和安全的同时，需要保证在其它车辆有非预期的行为(如突然刹车)时能保证乘客绝对的安全。
为了处理各种可能出现的情况，需要设计较复杂的速度规划和控制模型。在人工设计和实现中，很容易漏掉某些少见的情况和因素。同时，从驾驶员驾驶的汽车上可以采集很丰富的成熟的驾驶数据。由于机器学习方法可以很方便的从数据中学习模型，机器学习方法越来越多地应用到速度规划和控制。另外，在现实世界中，每个驾驶员都有不同的驾驶习惯和对安全舒适程度的界定，因此如果使用同一种规划和控制方法很难满足各种不同的需求，而机器学习方法可以很好的适配这种个性化驾驶习惯需求。
发明内容
本发明的发明人通过长期研究，认识到应用机器学习对车辆进行速度规划存在一些问题。
首先，通过机器学习方法学习到的模型的效果和训练数据直接相关。由于驾驶员通常都是在一个很舒适的范围内驾驶，收集到的训练数据很难覆盖到所有可能出现的场景(如车速极快而据前车距离很短等极端场景)，虽然泛化等技术可以一定程度上解决这个问题，但不能完全解决这个问题。另外，收集到的驾驶员的行为不一定完全满足舒适安全的需求，如有些驾驶员并不 能保证和前车的距离，因此当前车突然刹车时，驾驶车辆很难保证不碰上前车的情况下完全刹车。显然，用这种数据训练出来的模型同样也不能处理这种情况。
机器学习存在的另一个问题是很难对其训练出来的模型做局部微调。在应用于自动驾驶的速度规划和控制中，通常需要针对某种特殊情况对模型进行局部调整。而对机器学习得到的模型，对学习到的每个参数的调整，都可能带来全局的不易控制的影响。通过添加更多的训练用例也可以修正模型，但是这种方法周期比较长，并且最后训练出来的模型也不能完全预知。
本发明的目的在于克服现有技术中的缺点与不足，提出一种全新的用于车辆自动驾驶的速度规划方法、装置及计算装置。
根据本发明的一个方面，提供一种用于车辆自动驾驶的速度规划方法，具体包括：
首先为机器学习步骤，采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控制相关的控制参数数值；
其次为分区决策表取得步骤，对于输入空间进行分区，以及基于所述获得的机器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表；
再次为实时决策步骤，实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果。
进一步地，速度规划方法还可以包括实时控制步骤，基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
进一步地，在实时控制中，在发现确定分区的决策结果不符合预期时，可以对该分区的分区决策结果进行调整。
进一步地，对该分区的分区决策结果进行调整，可以依经验对该分区的分区决策结果进行调整，也可以通过机器学习方法对该分区进行学习，调整该分区的分区决策结果。
进一步地，各维特征分量可以包括当前车速、和前车的距离、前车相对 速度和最大车速。
进一步地，在机器学习步骤之前，需要对特征空间进行离散化编码。
进一步地，根据本发明实施例的速度规划方法，离散化编码方法优选为Tiling Coding，其中用只有一个Tiling的Tiling Coding对每个特征分量进行编码，并且将每个特征分量所处的维度划分成优选的7-13个区间，从而对所述输入空间进行分区。
进一步地，在分区决策表取得步骤中，首先需要计算采用离散化编码方法得到的离散化编码结果的空间大小，当所述空间大于确定阈值时，采用动态存储方法存储分区决策表，仅遍历训练空间的输入，存储相应决策模型的输出结果，同时除了分区决策表，还存储训练出的决策模型备用；当所述空间小于所述确定阈值时，采用静态存储方法存储分区决策表，遍历所有编码空间，存储决策模型的输出结果。
进一步地，在实时决策步骤中，首先采用所述离散化编码方法对所述输入特征量进行离散化编码，然后将获得的离散化编码结果作为分区决策表的索引，当分区决策表为采用静态存储方法存储时，直接获取分区决策表中已存储的决策结果；
进一步地，在实时决策步骤中，首先采用所述离散化编码方法对所述输入特征量进行离散化编码，然后将获得的离散化编码结果作为分区决策表的索引，当分区决策表为采用动态存储方法存储时，如果分区决策表内存储有该离散化编码结果的决策结果，则直接从分区决策表中获取决策结果；反之如果分区决策表内没有存储该离散化编码结果的决策结果，则调用所述存储的决策模型来得到决策结果，并将得到的决策结果加入到分区决策表。
进一步地，机器学习的方法可以采用监督式学习方法或非监督式学习方法，也可以采用增强式学习方法。
进一步地，对特征空间进行离散化编码时，只要两个输入最终的离散编码相同，即认为这两个输入属于同一个分区。
进一步地，离散化编码方法可以包括各类Coarse Coding方法，如Tile Coding。
进一步地，在实时控制中，在对该分区的分区决策结果进行调整后需要确定被调整过分区决策结果的分区的数目，当所述分区的数目超过预定阈值时，应当重新执行所述机器学习步骤和分区决策表取得步骤。
根据本发明的另一个方面，提供一种用于车辆自动驾驶的速度规划装置，该装置包括机器学习单元、分区决策表取得单元、实时决策单元。可选地，速度规划装置还可以包括实时控制单元。
其中，机器学习单元配置为采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控制相关的控制参数数值。
其中，分区决策表取得单元配置为对于输入空间进行分区，以及基于所述获得的机器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表。
其中，实时决策单元配置为实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果。
其中，实时控制单元配置为基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
进一步地，其中的实时控制单元还配置为在实时控制中，在发现确定分区的决策结果不符合预期时，对该分区的分区决策结果进行调整。
进一步地，其中的分区决策表取得单元还配置为对该分区的分区决策内容进行调整，可以依经验对该分区的分区决策内容进行调整，也可以通过机器学习方法对该分区进行学习，调整该分区的分区决策结果。
进一步地，其中实时决策单元还配置为各维特征分量包括当前车速、和前车的距离、前车相对速度和最大车速。
进一步地，本发明的车辆自动驾驶的速度规划装置还包括离散化编码单元，配置为对训练样本和所述实时决策阶段的输入特征量进行离散化编码。
进一步地，根据本发明实施例的速度规划装置，离散化编码单元采用的离散化编码方法为Tiling Coding，其中用只有一个Tiling的Tiling Coding对每个特征分量进行编码，并且将每个特征分量所处的维度划分成7-13个区间，从而对所述输入空间进行分区。
进一步地，分区决策表取得单元配置为计算采用离散化编码方法得到的离散化编码结果的空间大小，当所述空间大于确定阈值时，采用动态存储方 法存储分区决策表，仅遍历训练空间的输入，存储相应决策模型的输出结果，同时还存储训练出的决策模型备用；当所述空间小于所述确定阈值时，采用静态存储方法存储分区决策表，遍历所有编码空间，存储决策模型的输出结果。
进一步地，实时决策单元配置为采用所述离散化编码方法对所述输入特征量进行离散化编码，将获得的离散化编码结果作为分区决策表的索引，当分区决策表为采用静态存储方法存储时，直接获取分区决策表中已存储的决策结果。
进一步地，实时决策单元配置为采用所述离散化编码方法对所述输入特征量进行离散化编码，将获得的离散化编码结果作为分区决策表的索引，当分区决策表为采用动态存储方法存储时，如果分区决策表内存储有该离散化编码结果的决策结果，则直接从分区决策表中获取决策结果；反之如果分区决策表内没有存储该离散化编码结果的决策结果，则调用所述存储的决策模型来得到决策结果，并将得到的决策结果加入到分区决策表。
进一步地，机器学习的方法可以为监督式学习方法或非监督式学习方法，也可以为增强式学习方法。
进一步地，分区决策表取得单元配置为只要两个输入最终的离散编码相同，即认为这两个输入属于同一个分区。
进一步地，离散化编码方法选自可以包括各类Coarse Coding方法，如Tile Coding。
进一步地，反馈单元配置为确定被调整过分区决策结果的分区的数目，当所述分区的数目超过预定阈值时，引发所述机器学习单元和分区决策表取得单元重新进行机器学习操作和分区决策表取得操作。
根据本发明的另一个方面，提供一种用于车辆自动驾驶的速度规划的计算装置，包括存储部件和处理器，存储部件中存储有计算机可执行指令集合，当所述计算机可执行指令集合被所述处理器执行时，执行下述步骤：机器学习步骤，采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控制相关的控制参数数值；分区决策表取得步骤，对于输入空间进行分区，以及基于所述获得的机 器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表；实时决策步骤，实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果；以及实时控制步骤，基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
本发明采用分区决策表技术提供的速度规划方法、装置及计算装置适用于车辆自动驾驶技术，很好的解决了不能对机器学习训练出的模型做局部调整的问题，可以很容易修改某个分区的决策，而完全不影响其它分区的决策结果，从而完成局部调整。同时，分区决策表的直观特性可以很好的帮助发现和解决机器学习过程中存在的问题。分区决策表可以加速决策过程，查分区决策表可以获得更快的决策速度。
附图说明
从下面结合附图对本发明实施例的详细描述中，本发明的这些和/或其它方面和优点将变得更加清楚并更容易理解，其中：
图1是根据本发明一个实施例的用于车辆自动驾驶的速度规划方法的总体流程图；
图2示意性地示出了机器学习训练和应用的操作过程和输入、输出的示例性示意图；
图3示出了在应用机器学习方法之前，进行离散化编码处理的机器学习训练和应用的操作过程和输入、输出的示例性示意图；
图4示出了Coarse Coding类别中的Tile Coding离散化编码方法。
图5示出了根据本发明一个实施例的、在采用离散化编码情况下的整个决策分区表的建立和优化过程。
图6示出了根据本发明一个实施例的实时决策步骤的实现方法的流程图。
图7示出了根据本发明实施例的含有实时控制步骤的速度规划控制方法的700的流程图。
图8示出了根据本发明实施例的用于车辆自动驾驶的速度规划装置800的结构框图。
具体实施方式
为了使本领域技术人员更好地理解本发明，下面结合附图和具体实施方式对本发明作进一步详细说明。
在进行详细说明之前，首先介绍一下本发明的总体思想，以便于本领域技术人员把握本发明。
如前所述，发明人经实现和分析发现，利用机器学习具有严重依赖于训练数据，由于训练数据的局限性使得训练出来的模型不能处理一些实际突发情况等问题，以及还具有难于做局部微调，整体调整周期长且效果难以预知的问题。为此，发明人提出了本发明，将机器学习模型的训练和应用分别处理：在训练时，我们按照通用的机器学习方法来进行训练；在应用这些训练出来的模型之前，将输入空间按分成多个分区，并且将模型中决策结果存入到不同分区，从而形成一个按分区组织的决策表，由此，应用模型的决策过程转化为一个查分区决策表的过程。这样，可以很容易修改某个分区的决策，而完全不影响其它分区的决策结果，从而完成局部调整。同时，分区决策表的直观特性可以很好的帮助发现和解决机器学习过程中存在的问题。分区决策表可以加速决策过程，查分区决策表可以获得更快的决策速度。对于实时性要求较高的自动驾驶速度规划来说，易于调整局部策略和具有迅捷的决策速度是非常重要的。
下面结合图1描述根据本发明实施例的车辆自动驾驶方法示例。图1示出了本发明一个实施例的用于车辆自动驾驶的速度规划方法的总体流程图。
在步骤S110中，执行机器学习步骤，采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控制相关的控制参数数值。在步骤S110完成后，前进到步骤S120。
在步骤S120中，执行分区决策表取得步骤，对于输入空间进行分区，以及基于所述获得的机器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表。
在步骤S130中，进行实时决策，实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果。
根据本发明实施例，将机器学习模型的训练和应用分别处理。在训练时 按照通用的事先收集的训练用例训练模型，具体的机器学习方法可以是需要标注数据的监督式学习，也可以是不需要标注数据的非监督式学习，也可以是增强学习等。和传统方法不同的是，在应用这些训练出来的模型之前，将输入空间按训练时的离散化编码方法划分成多个分区，并且将模型中决策结果存入到不同分区，从而形成一个按分区组织的决策表。因此，采用本发明实施例的车辆自动驾驶的速度规划方法，就是将应用模型的决策过程转化为一个查分区决策表的过程。
为便于理解，图2示意性地示出了机器学习训练和应用的操作过程和输入、输出的示例性示意图。
在图2所示，针对输入的各个训练用例(训练样本)210，进行机器学习220，得到决策模型230，对输入空间分区240，结合分区和决策模型，得到分区决策表250。
在利用分区决策表进行决策时，对于到来的输入270，首先对其进行分区240，在确定所属分区后，到分区决策表250中查询获得对应的决策结果。作为进行分区的例子，对于连续维度的分区，例如可以基于车辆控制专家的经验知识来进行分区。
在一个示例中，在应用机器学习方法前，采用某种编码方式对输入空间进行离散化处理提取特征后再进行处理，这些离散化编码方法很自然的将连续输入空间划分成多个分区。这里的分区不限定是同一维度数据连续空间的分区，只要其最终的编码相同，即可以认为两个样本属于同一分区；换句话说这里的分区并不一定是按原空间连续划分的，而是按照离散编码的结果划分。图3示出了在应用机器学习方法之前，进行离散化编码处理的机器学习训练和应用的操作过程和输入、输出的示例性示意图，其中任何输入，不管是训练示例，还是应用示例，都要先进行离散化编码处理380。
图4示出了示出了粗编码(Coarse Coding)类别中的Tile Coding离散化编码方法。在图4所示的二维输入空间中，三个不同位置的Tiling将整个空间划分为不同的小区域，这些小区域即可作为分区。
作为输入空间的示例，可以由当前车速、和前车的距离、前车相对速度和最大车速四个维度构成。需要说明的是，这里的前车是一个宽泛的概念，并不局限于车辆。当前面没有任何物体时，可以虚拟出一辆前车，并设置虚拟前车的距离和相对速度。最大车速是由各种条件限制(如道路限制、天气 条线限制)下的最高行驶速度。因此，自动驾驶领域中的速度规划和控制的编码空间通常不大，因此比较适合进行分区。
发明人实践得到的一个优选例子为：用只有一个Tiling的Tiling Coding对每个输入维度进行编码，将每个输入维度划分成10个左右分区即可很好的满足自动驾驶的速度决策需求。比如对于最大车速，我们可以以10公里每小时的分区大小对其进行分区，结合现实情况，将其划分成[0，10，20，30，40，50，60，70，80，90，100，110，120]共12个分区。此时的分区数为104左右。对计算机而言，这样大小的分区表非常容易存储和处理。在某些极端情况下(如输入维度很多或离散化编码输出空间很大)，最后出来的编码空间可能超出计算机的存储空间。此时，基于来自现实世界的真实输入非常稀疏，可以采用哈希表或其它节省存储空间的方法来动态存储和处理分区决策表。
下面参考图5描述根据本发明一个实施例的、在采用离散化编码情况下的整个决策分区表的建立和优化过程示例，该过程可以用于执行图1所示的步骤S120。需要说明的是，图5中的阈值可以根据机器的存储空间设定。需要说明的是，此仅为示例，并非在机器学习训练之前，必须要对训练样本进行离散化编码，
如图5所示，在步骤S510中，计算机器学习方法采用离散化编码结果的空间大小。
在步骤S520中，判断空间大小是否大于某个阈值。
当步骤S520中判断结果为否时，前进到步骤S530，静态存储分区决策表：遍历所有编码空间，存储决策模型的输出结果；反之，当判断结果为是时，前进到步骤S540，用哈希或其它动态存储方法存储分区决策表：仅遍历训练空间的输入，存储相应决策模型的输出结果；接下来前进到步骤S550，除了分区决策表，还存储训练出的决策模型备用。需要说明的是，当需要动态存储分区决策表时，为了降低后续动态应用时动态建表的开销，可以选择提前添加训练集对应的表项，也可以选择添加其它输入集对应的表项，或者不添加任何表项，等后续需要时再动态添加。此时，由于后面需要动态***决策表项，除了初始分区决策表以外，也需要保存机器学习得到的决策模型，如步骤S550所示。
下面参考附图6描述根据本发明一个实施例的实时决策步骤S130的实现 方法的操作过程示例，这里，假设采用离散化编码方法对输入进行处理，而且分区决策表是通过图5所示的方法构建和存储的。
如图6所示，在步骤S610中，应用离散编码方法对输入进行编码。这里的输入可以为车辆自动驾驶过程中实时获得的当前车速、和前车的距离、前车相对速度和最大车速。不过此仅为示例，输入的维度和具体特征量可以根据速度规划方法的不同而不同。
在步骤S620中，将获得的离散化结果作为分区块的索引，查询分区决策表。
在分区决策表为采用静态存储方法存储的或者虽为采用动态存储方法存储的但是分区决策表内存储有该离散化编码结果的决策结果时，能够直接从分区决策表中获取决策结果；反之当分区决策表为采用动态存储方法存储且分区决策表内没有存储该离散化编码结果的决策结果时，则不能直接从分区决策表中获得决策结果。
在步骤S630中，判断是否得到决策结果。如果答案为“是”，则前进到步骤S650，返回决策结果。反之，答案为“否”，则前进到步骤S640，调用所述存储的决策模型来得到决策结果，并将得到的决策结果加入到分区决策表，然后前进到步骤S650。
需要说明的是，图6是在进行了离散编码而且先前视情况采用了静态存储或动态存储情况下的决策过程示例。不过这仅为示例，而不是作为本发明的限制，在不进行离散编码的情况下，利用分区决策表进行决策的过程可以不同，例如，此时可以直接根据输入来确定其所属分区，然后查询分区决策表，获得决策结果；在没有通过查询分区决策表得到决策结果的情况下，在存储了决策模型的情况下，可以调用存储的决策模型来得到决策结果，或者也可以直接给出反馈表明此时无法给出决策结果等等。
根据本发明实施例的速度规划方法，当发现确定分区的决策结果不符合预期时，能够迅速对该分区的分区决策结果进行调整。不需要修改训练用例集或调整训练参数重新训练，也不需要调整训练出来的模型的参数，而只需要修改分区决策表上相应分区的决策结果，从而保证所做的调整只局限于这个分区，而不会影响其它分区上的决策结果。
针对分区决策结果进行调整的具体方法可以是依经验对该分区的分区决策结果进行调整，也可以通过机器学习方法对该分区进行学习。比如，当发 现在当前车速是10公里每小时、和前车的距离是200米、前车相对速度是5公里每小时、最大车速是30公里每小时的时候速度控制不正常时，即可以参考某个有经验的驾驶员在这种情况下的反应，将其转换成相应的决策结果，存储到该输入所属区间对应的决策表项。或者也可以收集很多驾驶员在驾驶时的数据，将所有和该输入空间相应的决策结果收集起来，通过归纳或者简单的机器学习方法得到一个期望的决策结果。
在一个示例中，在获得决策结果后，执行实时控制步骤，基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。图7示出了根据本发明实施例的含有实时控制步骤的速度规划控制方法的700的流程图，图7中的步骤S710-S730与图1所示的步骤S110-S130类似，不同在于多了实时控制步骤S740，即基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
需要说明的是，这里根据速度规划决策结果，执行实时控制，并不是排他式的控制，而是可以结合自动驾驶的其他控制策略(例如转向控制等)来一起综合对车辆进行控制。
根据本发明一个优选实施例，当发现某些分区的决策结果不符合预期时，在对该分区的分区决策结果进行调整后可以确定被调整过分区决策结果的分区的数目，当所述分区的数目超过预定阈值时，则重新执行整体训练步骤，重新机器学习，然后得到新的决策模型，再次进行分区，即可以基于这些反馈结果来重新执行例如图1所示的步骤S110至步骤S130，即机器学习步骤和分区决策表取得步骤，在重新进行机器学习时，可以基于应用决策结果的反馈来尝试通过修改训练参数、添加相应的训练用例等方法来得到更好的机器学习。
根据本发明的另一实施例，还提供了一种用于车辆自动驾驶的速度规划装置。下面将结合图8进行说明，速度规划装置包括以下单元：机器学习单元810、分区决策表取得单元820和实时决策单元830。可选地，还可以包括实时控制单元830。
机器学习单元810配置为采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控 制相关的控制参数数值。
分区决策表取得单820配置为对于输入空间进行分区，以及基于所述获得的机器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表。
实时决策单元830配置为实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果。
实时控制单元840配置为基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
在一个示例中，速度规划装置还可以包括局部分区决策调整单元，配置为当发现确定分区的决策结果不符合预期时，对该分区的分区决策结果进行调整。
局部分区决策调整单元对该分区的分区决策内容进行调整时，可以依经验对该分区的分区决策内容进行调整，也可以通过机器学习方法对该分区进行学习，调整该分区的分区决策结果。
在一个示例中，速度规划装置还包括离散化编码单元，配置为对训练样本和所述实时决策阶段的输入特征量进行离散化编码。
因为最终的编码空间通常不大，比如在实现自动巡航***时的速度规划和控制中，输入只有当前车速、和前车的距离、前车相对速度和最大车速四个维度。离散化编码单元优选采用的离散化编码方法为Tiling Coding，只有一个Tiling的Tiling Coding对每个输入维度进行编码，将每个输入维度划分成优选的7-13个区间，这样即可很好的满足自动驾驶的速度决策需求。
因为在某些极端情况下(如输入维度很多或离散化编码输出空间很大)，最后出来的编码空间可能超出计算机的存储空间，所以分区决策表取得单元820还配置为计算采用离散化编码方法得到的离散化编码结果的空间大小。
当分区决策表取得单元820计算出采用离散化编码方法得到的离散化编码结果的空间大于所确定的阈值时，则决定选择采用动态存储方法存储分区决策表，此时仅遍历训练空间的输入，存储相应决策模型的输出结果，同时还存储训练出的决策模型备用。
相应的，本发明的实时决策单元830配置为先采用所述离散化编码方法对所述输入特征量进行离散化编码，再将获得的离散化编码结果作为分区决 策表的索引，因为此时分区决策表为采用动态存储方法存储，如果分区决策表内存储有该离散化编码结果的决策结果，则直接从分区决策表中获取决策结果；反之如果分区决策表内没有存储该离散化编码结果的决策结果，则调用所述存储的决策模型来得到决策结果，并将得到的决策结果加入到分区决策表。
当本发明的分区决策表取得单元820计算出采用离散化编码方法得到的离散化编码结果的空间小于所述确定阈值时，采用静态存储方法存储分区决策表，遍历所有编码空间，存储决策模型的输出结果。
相应的，本发明的实时决策单元830配置为先采用所述离散化编码方法对所述输入特征量进行离散化编码，再将获得的离散化编码结果作为分区决策表的索引，因为此时分区决策表为采用静态存储方法存储，可以直接获取分区决策表中已存储的决策结果。
机器学习单元810配置为采用的机器学习的方法可以为监督式学习方法或非监督式学习方法，也可以为增强式学习方法。
分区决策表取得单元820配置为只要两个输入最终的离散编码相同，即认为这两个输入属于同一个分区。
作为离散化编码单元配置为所采用的离散化编码方法可以选自各类Coarse Coding方法，如Tile Coding。
在一个示例中，速度规划装置800还可以包括反馈单元，配置为当发现某些分区的决策结果不符合预期时，在对该分区的分区决策结果进行调整后需要确定被调整过分区决策结果的分区的数目，当所述分区的数目超过预定阈值时，引发所述机器学习单元810和分区决策表取得单元820重新进行机器学习操作和分区决策表取得操作。
本发明另一个实施例还提供了一种用于车辆自动驾驶的速度规划的计算装置，包括存储部件和处理器，存储部件中存储有计算机可执行指令集合，当所述计算机可执行指令集合被所述处理器执行时，执行下述步骤：机器学习步骤，采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控制相关的控制参数数值；分区决策表取得步骤，对于输入空间进行分区，以及基于所述获得的机 器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表；实时决策步骤，实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果。
可选地，当所述计算机可执行指令集合被所述处理器执行时，还可以执行下述步骤：实时控制步骤，基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
需要说明的是，本文中的车辆应该做广义理解，包括各种大中小型车辆，也包括水上交通工具等。
需要说明的是，有关方法的各个步骤可以在车辆本地执行，也可以在云端执行，或者在本地和云端结合起来执行，有关数据的存储也可以存储在本地，或者存储在云端，或者在本地和云端结合起来存储。
以上已经描述了本发明的各实施例，上述说明是示例上性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。因此，本发明的保护范围应该以权利要求的保护范围为准。

Claims (23)

  1. 一种用于车辆自动驾驶的速度规划方法，包括：
    机器学习步骤，采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控制相关的控制参数数值；
    分区决策表取得步骤，对于输入空间进行分区，以及基于所述获得的机器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表；
    实时决策步骤，实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果。
  2. 根据权利要求1的速度规划方法，还包括：
    实时控制步骤，基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
  3. 根据权利要求1的方法，还包括：
    在发现确定分区的决策结果不符合预期时，对该分区的分区决策结果进行调整。
  4. 根据权利要求3的方法，所述对该分区的分区决策结果进行调整包括：
    依经验对该分区的分区决策结果进行调整；和/或
    通过机器学习方法对该分区进行学习，调整该分区的分区决策结果。
  5. 根据权利要求1的方法，所述各维特征分量包括：
    当前车速、和前车的距离、前车相对速度和最大车速。
  6. 根据权利要求1的方法，还包括，在机器学习步骤之前，对特征空间进行离散化编码。
  7. 根据权利要求6所述的方法，其特征在于，所述分区决策表取得步骤包括：
    计算采用离散化编码方法得到的离散化编码结果的空间大小：
    当所述空间大于确定阈值时，采用动态存储方法存储分区决策表，仅遍历训练空间的输入，存储相应决策模型的输出结果，同时除了分区决策表，还存储训练出的决策模型备用；以及
    当所述空间小于所述确定阈值时，采用静态存储方法存储分区决策表，遍历所有编码空间，存储决策模型的输出结果。
  8. 根据权利要求7所述的方法，其特征在于，所述实时决策步骤包括：
    采用所述离散化编码方法对所述输入特征量进行离散化编码；
    将获得的离散化编码结果作为分区决策表的索引，当分区决策表为采用静态存储方法存储时，直接获取分区决策表中已存储的决策结果；
    将获得的离散化编码结果作为分区决策表的索引，当分区决策表为采用动态存储方法存储时，如果分区决策表内存储有该离散化编码结果的决策结果，则直接从分区决策表中获取决策结果；反之如果分区决策表内没有存储该离散化编码结果的决策结果，则调用所述存储的决策模型来得到决策结果，并将得到的决策结果加入到分区决策表。
  9. 根据权利要求6所述的方法，其特征在于，只要两个输入最终的离散编码相同，即认为这两个输入属于同一个分区。
  10. 根据权利要求6所述的方法，其特征在于，所述离散化编码方法为粗编码方法中的一种。
  11. 根据权利要求2所述的方法，还包括：
    确定被调整过分区决策结果的分区的数目，当所述分区的数目超过预定阈值时，重新执行所述机器学习步骤和分区决策表取得步骤。
  12. 一种用于车辆自动驾驶的速度规划装置，包括：
    机器学习单元，配置为采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控制相关的控制参数数值；
    分区决策表取得单元，配置为对于输入空间进行分区，以及基于所述获得的机器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表；
    实时决策单元，配置为实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果。
  13. 根据权利要求12的速度规划装置，还包括：
    实时控制单元，配置为基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
  14. 根据权利要求12的速度规划装置，还包括：
    局部分区决策调整单元，在发现确定分区的决策结果不符合预期时，对该分区的分区决策结果进行调整。
  15. 根据权利要求14的速度规划装置，所述对该分区的分区决策内容进行调整包括：
    依经验对该分区的分区决策内容进行调整；和/或
    通过机器学习方法对该分区进行学习，调整该分区的分区决策结果。
  16. 根据权利要求12的速度规划装置，所述各维特征分量包括：
    当前车速、和前车的距离、前车相对速度和最大车速。
  17. 根据权利要求12的速度规划装置，还包括：
    离散化编码单元，配置为对训练样本和所述实时决策阶段的输入特征量进行离散化编码。
  18. 根据权利要求17所述的速度规划装置，其特征在于，所述分区决策表取得单元配置为：
    计算采用离散化编码方法得到的离散化编码结果的空间大小：
    当所述空间大于确定阈值时，采用动态存储方法存储分区决策表，仅遍历训练空间的输入，存储相应决策模型的输出结果，同时还存储训练出的决策模型备用；以及
    当所述空间小于所述确定阈值时，采用静态存储方法存储分区决策表，遍历所有编码空间，存储决策模型的输出结果。
  19. 根据权利要求18所述的速度规划装置，其特征在于，实时决策单元配置为：
    采用所述离散化编码方法对所述输入特征量进行离散化编码；
    将获得的离散化编码结果作为分区决策表的索引，当分区决策表为采用静态存储方法存储时，直接获取分区决策表中已存储的决策结果；
    将获得的离散化编码结果作为分区决策表的索引，当分区决策表为采用动态存储方法存储时，如果分区决策表内存储有该离散化编码结果的决策结果，则直接从分区决策表中获取决策结果；反之如果分区决策表内没有存储该离散化编码结果的决策结果，则调用所述存储的决策模型来得到决策结果，并将得到的决策结果加入到分区决策表。
  20. 根据权利要求17所述的速度规划装置，其特征在于，只要两个输入最终的离散编码相同，即认为这两个输入属于同一个分区。
  21. 根据权利要求17所述的速度规划装置，其特征在于，所述离散化编码方法为粗编码方法中的一种。
  22. 根据权利要求12所述的速度规划装置，还包括：
    反馈单元，配置为确定被调整过分区决策结果的分区的数目，当所述分区的数目超过预定阈值时，引发所述机器学习单元和分区决策表取得单元重新进行机器学习操作和分区决策表取得操作。
  23. 一种用于车辆自动驾驶的速度规划的计算装置，包括存储部件和处理器，存储部件中存储有计算机可执行指令集合，当所述计算机可执行指令集合被所述处理器执行时，执行下述步骤：
    机器学习步骤，采用训练样本集进行机器学习，获得机器学习模型，每个训练样本由形成输入空间的多维特征分量和形成输出空间的决策结果来描述，所述多维特征分量的每维是用于描述车辆特定时刻状态的、与速度规划有关的变量，所述决策结果指示下一时刻的预期速度和/或与速度控制相关的控制参数数值；
    分区决策表取得步骤，对于输入空间进行分区，以及基于所述获得的机器学习模型，得到与确定分区对应的决策结果，形成各个分区对应于相应决策结果的分区决策表；
    实时决策步骤，实时获得行驶中车辆的各维特征分量作为输入特征量，确定该输入特征量所属于的输入分区，基于所确定的分区，查询分区决策表来获得相应的决策结果；以及
    实时控制步骤，基于所获得的决策结果，对车辆发出控制命令，从而控制车辆的速度。
PCT/CN2016/085310 2016-06-08 2016-06-08 车辆自动驾驶的速度规划方法、装置及计算装置 WO2017210901A1 (zh)

Priority Applications (6)

Application Number Priority Date Filing Date Title
PCT/CN2016/085310 WO2017210901A1 (zh) 2016-06-08 2016-06-08 车辆自动驾驶的速度规划方法、装置及计算装置
CN201680001425.1A CN107182206B (zh) 2016-06-08 2016-06-08 车辆自动驾驶的速度规划方法、装置及计算装置
ES16904354T ES2833674T3 (es) 2016-06-08 2016-06-08 Método y aparato de planificación de velocidad y aparato de cálculo para la conducción automática de un vehículo
EP16904354.4A EP3460613B1 (en) 2016-06-08 2016-06-08 Speed planning method and apparatus and calculating apparatus for automatic driving of vehicle
US16/308,353 US10564644B2 (en) 2016-06-08 2016-06-08 Speed planning method and apparatus and calculating apparatus for automatic driving of vehicle
US16/735,960 US11747818B2 (en) 2016-06-08 2020-01-07 Speed planning method and apparatus and calculating apparatus for automatic driving of vehicle

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/085310 WO2017210901A1 (zh) 2016-06-08 2016-06-08 车辆自动驾驶的速度规划方法、装置及计算装置

Related Child Applications (2)

Application Number Title Priority Date Filing Date
US16/308,353 A-371-Of-International US10564644B2 (en) 2016-06-08 2016-06-08 Speed planning method and apparatus and calculating apparatus for automatic driving of vehicle
US16/735,960 Continuation US11747818B2 (en) 2016-06-08 2020-01-07 Speed planning method and apparatus and calculating apparatus for automatic driving of vehicle

Publications (1)

Publication Number Publication Date
WO2017210901A1 true WO2017210901A1 (zh) 2017-12-14

Family

ID=59830175

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/085310 WO2017210901A1 (zh) 2016-06-08 2016-06-08 车辆自动驾驶的速度规划方法、装置及计算装置

Country Status (5)

Country Link
US (2) US10564644B2 (zh)
EP (1) EP3460613B1 (zh)
CN (1) CN107182206B (zh)
ES (1) ES2833674T3 (zh)
WO (1) WO2017210901A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110196593A (zh) * 2019-05-16 2019-09-03 济南浪潮高新科技投资发展有限公司 一种自动驾驶多场景环境检测及决策***及方法
AU2018286585A1 (en) * 2018-12-18 2020-07-02 Beijing Voyager Technology Co., Ltd. Systems and methods for autonomous driving
AU2018286610A1 (en) * 2018-12-18 2020-07-02 Beijing Voyager Technology Co., Ltd. Systems and methods for determining driving action in autonomous driving
KR20210101172A (ko) * 2020-11-06 2021-08-18 바이두 온라인 네트웍 테크놀러지 (베이징) 캄파니 리미티드 자율 주행을 위한 속도 계획 방법, 장치, 설비, 매체 및 차량

Families Citing this family (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109901446B (zh) 2017-12-08 2020-07-07 广州汽车集团股份有限公司 路口通行控制方法、装置及***
US20190185012A1 (en) 2017-12-18 2019-06-20 PlusAI Corp Method and system for personalized motion planning in autonomous driving vehicles
US11273836B2 (en) * 2017-12-18 2022-03-15 Plusai, Inc. Method and system for human-like driving lane planning in autonomous driving vehicles
US11130497B2 (en) 2017-12-18 2021-09-28 Plusai Limited Method and system for ensemble vehicle control prediction in autonomous driving vehicles
CN108227710B (zh) * 2017-12-29 2022-10-04 商汤集团有限公司 自动驾驶控制方法和装置、电子设备、程序和介质
US11126199B2 (en) * 2018-04-16 2021-09-21 Baidu Usa Llc Learning based speed planner for autonomous driving vehicles
DE102018206743A1 (de) * 2018-05-02 2019-11-07 Bayerische Motoren Werke Aktiengesellschaft Verfahren zum Betreiben eines Fahrerassistenzsystems eines Egofahrzeugs mit wenigstens einem Umfeldsensor zum Erfassen eines Umfelds des Egofahrzeugs, Computer-lesbares Medium, System, und Fahrzeug
US11093493B1 (en) * 2018-09-28 2021-08-17 Amazon Technologies, Inc. Dynamically switching between query and scan for optimizing table reads
CN110376583B (zh) * 2018-09-30 2021-11-19 毫末智行科技有限公司 用于车辆传感器的数据融合方法及装置
CN110968088B (zh) * 2018-09-30 2023-09-12 百度（美国）有限责任公司 车辆控制参数的确定方法、装置、车载控制器和无人车
CN111123735B (zh) * 2018-10-31 2022-09-09 百度在线网络技术（北京）有限公司 自动驾驶仿真运行方法和装置
US10955853B2 (en) 2018-12-18 2021-03-23 Beijing Voyager Technology Co., Ltd. Systems and methods for autonomous driving
CN110281949B (zh) * 2019-06-28 2020-12-18 清华大学 一种自动驾驶统一分层决策方法
CN110673602B (zh) * 2019-10-24 2022-11-25 驭势科技（北京）有限公司 一种强化学习模型、车辆自动驾驶决策的方法和车载设备
CN111267109B (zh) * 2020-03-24 2021-07-02 华中科技大学 一种基于强化学习的机器人速度规划方法和***
TW202137078A (zh) * 2020-03-24 2021-10-01 廣達電腦股份有限公司 資料處理系統及資料處理方法
WO2021203385A1 (zh) * 2020-04-09 2021-10-14 华为技术有限公司 一种生成指令单元组的方法及装置
CN112766310B (zh) * 2020-12-30 2022-09-23 嬴彻星创智能科技(上海)有限公司 一种节油换道决策方法和***
CN113377112B (zh) * 2021-06-30 2023-05-05 东风商用车有限公司 一种自动驾驶速度规划及状态协调方法和装置
CN117519206A (zh) * 2023-12-07 2024-02-06 北京百度网讯科技有限公司 基于生成式扩散模型的自动驾驶模型、方法、装置和车辆

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120046819A1 (en) * 2008-09-25 2012-02-23 Honeywell International Inc. System for integrating dynamically observed and static information for route planning in a graph based planner
CN103383265A (zh) * 2013-07-04 2013-11-06 浙江吉利汽车研究院有限公司杭州分公司 汽车自动驾驶***
CN103777631A (zh) * 2013-12-16 2014-05-07 北京交控科技有限公司 一种自动驾驶的控制***及方法
CN105185141A (zh) * 2015-10-14 2015-12-23 浙江大学 一种基于引力场的车辆自动驾驶引导方法
CN105189237A (zh) * 2013-03-19 2015-12-23 马西夫分析有限公司 用于控制自动驾驶或部分自动驾驶的陆地车辆的设备
CN105487537A (zh) * 2015-11-06 2016-04-13 福州华鹰重工机械有限公司 一种车辆运动规划方法和无人车

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8135531B2 (en) * 2002-06-12 2012-03-13 Nmhg Oregon, Llc Predictive vehicle controller
US7953521B2 (en) * 2005-12-30 2011-05-31 Microsoft Corporation Learning controller for vehicle control
US8612107B2 (en) * 2008-06-10 2013-12-17 The Regents Of The University Of Michigan Method, control apparatus and powertrain system controller for real-time, self-learning control based on individual operating style
WO2010106587A1 (ja) * 2009-03-18 2010-09-23 パナソニック株式会社 ニューラルネットワークシステム
CN103035121A (zh) * 2012-12-06 2013-04-10 南京航空航天大学 一种智能车辆自主行驶动态轨迹规划方法及***
US9081651B2 (en) * 2013-03-13 2015-07-14 Ford Global Technologies, Llc Route navigation with optimal speed profile
US20150046060A1 (en) * 2013-08-12 2015-02-12 Mitsubishi Electric Research Laboratories, Inc. Method and System for Adjusting Vehicle Settings
US9346167B2 (en) * 2014-04-29 2016-05-24 Brain Corporation Trainable convolutional network apparatus and methods for operating a robotic vehicle
US10133250B2 (en) * 2014-06-20 2018-11-20 Veritone Alpha, Inc. Managing construction of decision modules to control target systems
US9766625B2 (en) * 2014-07-25 2017-09-19 Here Global B.V. Personalized driving of autonomously driven vehicles
US9630318B2 (en) * 2014-10-02 2017-04-25 Brain Corporation Feature detection apparatus and methods for training of robotic navigation
US10026506B1 (en) * 2015-02-06 2018-07-17 Brain Trust Innovations I, Llc System, RFID chip, server and method for capturing vehicle data
US9864918B2 (en) * 2015-11-04 2018-01-09 Ford Global Technologies, Llc Predicting vehicle movements based on driver body language

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120046819A1 (en) * 2008-09-25 2012-02-23 Honeywell International Inc. System for integrating dynamically observed and static information for route planning in a graph based planner
CN105189237A (zh) * 2013-03-19 2015-12-23 马西夫分析有限公司 用于控制自动驾驶或部分自动驾驶的陆地车辆的设备
CN103383265A (zh) * 2013-07-04 2013-11-06 浙江吉利汽车研究院有限公司杭州分公司 汽车自动驾驶***
CN103777631A (zh) * 2013-12-16 2014-05-07 北京交控科技有限公司 一种自动驾驶的控制***及方法
CN105185141A (zh) * 2015-10-14 2015-12-23 浙江大学 一种基于引力场的车辆自动驾驶引导方法
CN105487537A (zh) * 2015-11-06 2016-04-13 福州华鹰重工机械有限公司 一种车辆运动规划方法和无人车

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
AU2018286585A1 (en) * 2018-12-18 2020-07-02 Beijing Voyager Technology Co., Ltd. Systems and methods for autonomous driving
AU2018286610A1 (en) * 2018-12-18 2020-07-02 Beijing Voyager Technology Co., Ltd. Systems and methods for determining driving action in autonomous driving
AU2018286585B2 (en) * 2018-12-18 2020-10-22 Beijing Voyager Technology Co., Ltd. Systems and methods for autonomous driving
AU2018286610B2 (en) * 2018-12-18 2020-10-29 Beijing Voyager Technology Co., Ltd. Systems and methods for determining driving action in autonomous driving
CN110196593A (zh) * 2019-05-16 2019-09-03 济南浪潮高新科技投资发展有限公司 一种自动驾驶多场景环境检测及决策***及方法
CN110196593B (zh) * 2019-05-16 2022-03-01 山东浪潮科学研究院有限公司 一种自动驾驶多场景环境检测及决策***及方法
KR20210101172A (ko) * 2020-11-06 2021-08-18 바이두 온라인 네트웍 테크놀러지 (베이징) 캄파니 리미티드 자율 주행을 위한 속도 계획 방법, 장치, 설비, 매체 및 차량
KR102566603B1 (ko) 2020-11-06 2023-08-14 바이두 온라인 네트웍 테크놀러지 (베이징) 캄파니 리미티드 자율 주행을 위한 속도 계획 방법, 장치, 설비, 매체 및 차량

Also Published As

Publication number Publication date
EP3460613A4 (en) 2019-06-05
US10564644B2 (en) 2020-02-18
EP3460613B1 (en) 2020-10-07
CN107182206B (zh) 2020-07-03
US20190265713A1 (en) 2019-08-29
ES2833674T3 (es) 2021-06-15
US11747818B2 (en) 2023-09-05
EP3460613A1 (en) 2019-03-27
CN107182206A (zh) 2017-09-19
US20200218275A1 (en) 2020-07-09

Similar Documents

Publication Publication Date Title
WO2017210901A1 (zh) 车辆自动驾驶的速度规划方法、装置及计算装置
CN106828493B (zh) 一种自动驾驶车辆分层式纵向规划控制***及方法
CN105243430B (zh) 列车节能运行的目标速度曲线的优化方法
WO2022148282A1 (zh) 一种车辆轨迹规划方法、装置、存储介质及设备
CN107229973B (zh) 一种用于车辆自动驾驶的策略网络模型的生成方法及装置
CN106154831B (zh) 一种基于学习法的智能汽车纵向神经滑模控制方法
WO2022227720A1 (zh) 车辆速度跟踪控制方法、装置、设备及存储介质
CN108860148A (zh) 基于驾驶员跟车特性安全距离模型的自适应巡航控制方法
US10320354B1 (en) Controlling a volume level based on a user profile
CN108639061B (zh) 一种自动驾驶车辆主动防晕车辅助驾驶控制方法
TW201238799A (en) Methods and systems of saving energy control
CN110254425A (zh) 一种模拟驾驶员踏板控制的定速巡航方法及***
CN108099908B (zh) 一种车辆自适应巡航优化控制计算方法
CN109204390B (zh) 一种基于深度学习的列车控制方法
CN106056238B (zh) 列车区间运行轨迹的规划方法
CN108845498B (zh) 一种考虑反应滞后时间的驾驶员跟驰模型
US20210341886A1 (en) System and Method of Efficient, Continuous, and Safe Learning Using First Principles and Constraints
WO2022252457A1 (zh) 一种自动驾驶控制方法、装置、设备及可读存储介质
CN113570855B (zh) 可变车道的控制方法、装置、设备以及存储介质
US20170355370A1 (en) Vehicle control system
CN110427690A (zh) 一种基于全局粒子群算法生成ato速度曲线的方法及装置
CN112009397A (zh) 一种自动驾驶路测数据分析方法及装置
CN113635900B (zh) 一种预测巡航过程中基于能量管理的换道决策控制方法
Hua A new car-following model considering recurrent neural network
Pan et al. Energy-optimized adaptive cruise control strategy design at intersection for electric vehicles based on speed planning

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: 16904354

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2016904354

Country of ref document: EP

Effective date: 20181217