Control method, device and equipment of vehicle and vehicle

Abstract

The application provides a vehicle control method, device and equipment, and a vehicle. The method comprises the following steps: a control device acquires vehicle information and environment information, and determines a first target point from a first strategy space according to the vehicle information and the environment information. The first strategy space comprises M*N target points determined by M spatial target points and N time target points. The spatial target point indicates a position that can be reached by the vehicle within a preset time or a preset distance. The time target point indicates a driving speed and / or acceleration of the vehicle within the preset time in a direction of a corresponding spatial target point. Then, a target strategy is determined according to the first target point, and the vehicle is controlled to drive according to the target strategy. The control device determines the target point that determines the target strategy in the strategy space with explicit semantics, realizes the driving control of the vehicle, avoids the mode collapse problem of the algorithm, and improves the accuracy of the driving decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent vehicles, and in particular to a vehicle control method, device, equipment and vehicle. BACKGROUND

[0002] In the field of automatic driving of an automobile, a vehicle needs to predict a driving path of a surrounding traffic participant according to dynamic and static environmental information and global navigation information around the vehicle, and then perform driving control on the vehicle according to the predicted driving path. Since there are many feature dimensions to be considered for making a driving decision, directly predicting the driving path based on the environmental information and the navigation information may cause some features in the environmental information and the navigation information to lack representation ability, that is, modal collapse of an algorithm occurs, thereby causing the driving decision result to be inaccurate. SUMMARY

[0003] The vehicle control method, device, equipment and vehicle provided by the embodiments of the present application are to improve the accuracy of driving decision in the implementation of automatic driving.

[0004] In a first aspect, the embodiments of the present application provide a vehicle control method, which comprises: obtaining vehicle information and environmental information, the vehicle information comprising at least one of driving intention information, driving trajectory information or pose information of a vehicle, and the environmental information comprising topological information of a road and / or information of a target object in the road; determining a first target point from a first strategy space according to the vehicle information and the environmental information, the coordinates of the first target point in the first strategy space comprising a value of a first space target point and a value of a first time target point, the first strategy space comprising M*N target points, the M*N target points being determined by M space target points and N time target points, the space target point indicating a position that the vehicle can reach within a preset time or a preset distance, and the time target point indicating a driving speed and / or acceleration of the vehicle within the preset time in a direction of a corresponding space target point, M and N being positive integers greater than 1; determining a target strategy according to the first target point; and controlling the vehicle to drive according to the target strategy.

[0005] The vehicle control method provided by the first aspect controls the device to determine a target point for determining a target strategy in a strategy space with explicit semantics based on vehicle information and environmental information, thereby achieving driving control of the vehicle, avoiding the problem of modal collapse of an algorithm, and improving the accuracy of driving decision.

[0006] In a possible implementation, the M space target points form a space surface in a Euclidean space, the space surface being perpendicular to a road direction and being at a first distance from the vehicle, the first distance being greater than or equal to a preset distance.

[0007] The control method of the vehicle provided by the embodiment gives a clear semantic to the spatial dimension of the first strategy space, so that the target strategy determined by the control device is more accurate.

[0008] In a possible implementation, the first distance is determined based on the preset time and the current speed of the vehicle.

[0009] The control method of the vehicle provided by the embodiment determines the position indicated by the M spatial target points based on the preset time and the current speed of the vehicle, so that the driving position indicated by the first target point is more in line with the driving rule of the vehicle, that is, the human nature of the target strategy is improved.

[0010] In a possible implementation, the first target point is determined from the first strategy space according to the vehicle information and the environment information, including: determining a first probability distribution according to the vehicle information and the environment information, the first probability distribution including probabilities of the M*N target points for determining the target strategy; and determining the first target point from the first strategy space according to the first probability distribution.

[0011] The control method of the vehicle provided by the embodiment predicts the probability of each target point for determining the target strategy by the control device, and determines the first target point according to the probability of each target point for determining the target strategy, so that the target strategy determined by the first target point can accurately control the driving of the vehicle.

[0012] Optionally, the control device takes the target point corresponding to the maximum value in the first probability distribution as the first target point, which can quickly select a global optimal solution and has high processing efficiency.

[0013] Optionally, the control device determines K candidate target points from the first strategy space according to the first probability distribution, and determines the first target point from the K candidate target points, K being an integer greater than 1. The K candidate target points respectively correspond to K probability values in the first probability distribution, the K probability values include a maximum value, or the K probability values include the maximum value and probability values within a preset range of the maximum value, or the K probability values are the first K probability values in the first probability distribution in descending order. In this case, the first target point is not limited to the target point with the highest probability, which increases the dimension of determining the first target point.

[0014] Optionally, the control device determines the first target point from the first strategy space according to the first probability distribution by using a decision tree algorithm, or determines the first target point from the first strategy space according to the first probability distribution by using a joint optimization algorithm. The combination of the first probability distribution and the white box algorithm improves the reliability of the target strategy.

[0015] In a possible implementation, the maximum value in the first probability distribution corresponds to a target point, and the target point is taken as the first target point. The method comprises the following steps: determining a predicted trajectory based on the target point corresponding to the maximum value in the first probability distribution; performing driving safety evaluation on the predicted trajectory; and taking the target point corresponding to the maximum value in the first probability distribution as the first target point when the driving safety evaluation is passed. In this way, the driving safety can be ensured when the vehicle is controlled according to the target strategy, and the reliability of the driving control is improved.

[0016] Optionally, the control device performs driving safety evaluation on the predicted trajectories corresponding to the K candidate target points, and takes the candidate target point with higher driving safety as the first target point. In this way, the driving safety can be ensured when the vehicle is controlled according to the target strategy.

[0017] Optionally, the control device determines the first target point from the K candidate target points by using a decision tree algorithm, or determines the first target point from the K candidate target points by using a joint optimization algorithm. The target point is further screened based on the first probability distribution by using the white-box algorithm, and the reliability of the target strategy determined based on the first target point is improved.

[0018] In a possible implementation, the first target point is determined from the K candidate target points by using a decision tree algorithm, and the method comprises the following steps: determining a path of each candidate target point in the decision tree; determining a generation value of the candidate target point according to the path of the candidate target point in the decision tree; and taking the candidate target point with the minimum generation value as the first target point. In this way, the target strategy that best meets the driving control requirement can be determined.

[0019] In a possible implementation, the decision tree comprises a first-level node, a second-level node, and a third-level node. The first-level node is a parent node of the second-level node, and the second-level node is a parent node of the third-level node. The first-level node represents a probability that a target point of the first strategy space is used to determine the target strategy. The second-level node represents a predicted trajectory of the vehicle. The third-level node represents an interaction label and / or a predicted trajectory of at least one target object. The interaction label is used to indicate an interaction strategy of the vehicle to the target object, and the interaction strategy comprises at least one of yielding, taking, and bypassing. By defining the first-level node representing the probability that the target point of the first strategy space is used to determine the target strategy, the second-level node representing the predicted trajectory, and the third-level node representing the interaction label and / or the predicted trajectory of the target object, the parameter complexity of the white-box algorithm is reduced compared with defining the first-level node representing the vehicle information and the road information.

[0020] In a possible implementation, the nodes of the first sub-level of the first level represent spatial target points of the first strategy space for determining probabilities of the target strategy, the nodes of the second sub-level of the first level represent time target points of the first strategy space for determining probabilities of the target strategy, and the nodes of the first sub-level of the first level are parent nodes of the nodes of the second sub-level of the first level. The spatial target points and the time target points are respectively arranged as different level nodes, so that the parameters of different levels in the decision tree have clear semantics, and the accuracy of the decision tree algorithm in determining the target strategy is improved.

[0021] In a possible implementation, the nodes of the first sub-level of the third level represent interaction labels of the at least one target object, the nodes of the second sub-level of the third level represent predicted trajectories of the at least one target object; or the nodes of the first sub-level of the third level represent interaction labels of a first target object in the at least one target object, the second sub-level of the third level represents a predicted trajectory of the first target object, the third sub-level of the third level represents interaction labels of a second target object in the at least one target object, and the fourth sub-level of the third level represents a predicted trajectory of the second target object, and so on; or the nodes of the first sub-level of the third level represent predicted trajectories of a first target object in the at least one target object, the nodes of the second sub-level of the third level represent predicted trajectories of a second target object in the at least one target object, and so on.

[0022] The control method of the vehicle provided by the implementation has the following advantages. On the one hand, when different target objects are in the same sub-level in the third level, the order between different target objects is not fixed, which facilitates obtaining an optimal solution, and when different target objects are in different sub-levels in the third level, the interaction relationship between different target objects is simplified in the decision tree algorithm, and the complexity of the algorithm is relatively low. On the other hand, when the third level does not include a sub-level of an interaction label, the constraint on the predicted trajectory in the decision tree is less, and the possibility of the predicted trajectory is more, and when the third level includes a sub-level of an interaction label, the dimension of the predicted trajectory that can be defined is less, and the search space of the decision tree is smaller, and the processing efficiency is relatively high.

[0023] In a possible implementation, the method further includes: determining, according to the vehicle information and the environment information, a predicted trajectory and an interaction label of the at least one target object, the interaction label being used to indicate an interaction strategy of the vehicle to the target object, the interaction strategy including at least one of yielding, cutting in, and detouring, and the predicted trajectory and the interaction label of the target object being used to perform driving safety evaluation.

[0024] In a possible implementation, the determining, according to the vehicle information and the environment information, of the predicted trajectory and the interaction label of the at least one target object comprises: determining, according to the vehicle information, the environment information and navigation information, the predicted trajectory and the interaction label of the at least one target object, the navigation information indicating a planned driving path of the vehicle.

[0025] In a possible implementation, the determining, according to the vehicle information and the environment information, of the first probability distribution comprises: determining, according to the vehicle information and the environment information, a probability distribution of the M spatial target points; determining, for a probability of each spatial target point in the M spatial target points, a conditional probability distribution of the N temporal target points under the spatial target point; and determining, according to the conditional probability distribution of the N temporal target points under the M spatial target points, the first probability distribution.

[0026] In a possible implementation, the method further comprises: determining, according to the vehicle information and the environment information, an initial strategy space; and discretizing the initial strategy space to obtain the M spatial target points and the N temporal target points.

[0027] In a possible implementation, the determining, according to the vehicle information and the environment information, of the first target point from the first strategy space comprises: inputting the vehicle information, the environment information and the M by N target points into a first model, and outputting, by the first model, at least one of the following: the first probability distribution; the predicted trajectory corresponding to each of the M by N target points; the interaction label of the at least one target object; the predicted trajectory of the at least one target object; wherein the first model is obtained by training a neural network model. The neural network model is used to predict the information related to the decision target strategy, and the human-likeness of the target strategy is improved.

[0028] In a possible implementation, the method further comprises: obtaining a plurality of training samples, the training samples comprising input data and expected output data, the expected output data being used to indicate an expected output of the neural network model when the input data is input, the input data comprising M' by N' target points of a second strategy space, vehicle information and environment information in a human driving interaction process, M' and N' being positive integers greater than 1, the expected output data comprising at least one of a spatial target point true value, a temporal target point true value, a target point true value, a vehicle trajectory true value, a target object trajectory true value and an interaction label true value; inputting the plurality of training samples into the neural network model, training the neural network model, and obtaining the first model after a loss value converges to a preset value, the loss value being determined based on the expected output data and a predicted value of the input data by the neural network model. The model is trained based on the vehicle information and the environment information in the human driving interaction process, so that the target strategy determined based on the first model has higher human-likeness.

[0029] In a possible implementation, the plurality of training samples are input into the neural network model, the neural network model is trained, and after the loss value converges to a preset value, the first model is obtained, including: inputting the plurality of training samples into the neural network model, and obtaining the first model through the following training process; the training process includes a first training stage and a second training stage; in the first training stage, the probabilities corresponding to the M' spatial target points are predicted by the neural network model, and the first loss value is determined according to the spatial target point with the highest probability in the M' spatial target points and the spatial target point true value; the first training stage is ended when the first loss value converges to a first preset value through iterative training; in the second training stage, the probabilities corresponding to the N' time target points are predicted by the neural network model under the spatial target point true value, and the second loss value is determined according to the time target point with the highest probability in the N' time target points and the time target point true value; the second training stage is ended when the second loss value converges to a second preset value through iterative training; or, the training process includes a third training stage; in the third training stage, the joint probability corresponding to the M'N' target points is predicted by the neural network model, and the third loss value is determined according to the target point with the highest joint probability in the M'N' target points and the target point true value; the third training stage is ended when the third loss value converges to a third preset value through iterative training.

[0030] In the above method, the model is trained based on the probabilities of the spatial target points and the probabilities of the time target points respectively, so that the interpretability of each step of inference of the model is higher, and the coupling characteristics of the probabilities of the spatial target points and the probabilities of the time target points are trained, which simplifies the training process and enables the model training to converge faster.

[0031] In a possible implementation, the training process further includes a fourth training stage: after the first training stage of the training process is ended, a trajectory corresponding to the spatial target point true value is predicted by the neural network model; or, after the second training stage of the training process is ended, a trajectory corresponding to the time target point true value is predicted by the neural network model; or, after the third training stage of the training process is ended, a trajectory corresponding to the target point true value is predicted by the neural network model; a fourth loss value is determined according to the trajectory and the trajectory true value, and the fourth training stage is ended when the fourth loss value converges to a fourth preset value through iterative training.

[0032] In a possible implementation, the training process further includes a fifth training stage and / or a sixth training stage; in the fifth training stage, a trajectory of the target object is predicted by the neural network model, a fifth loss value is determined according to the trajectory of the target object and a trajectory ground truth of the target object, and the fifth training stage is ended when the fifth loss value converges to a fifth preset value through iterative training; in the sixth training stage, an interaction label of the target object is predicted by the neural network model, a sixth loss value is determined according to the interaction label of the target object and an interaction label ground truth, and the sixth training stage is ended when the sixth loss value converges to a sixth preset value through iterative training.

[0033] In a second aspect, the embodiments of the present application provide a control device, including modules for performing the method in the first aspect or the possible implementation manners of the first aspect.

[0034] In a third aspect, the embodiments of the present application provide a chip, including a processor configured to invoke and run computer instructions from a memory, so that a device installed with the chip performs the method in the first aspect or the possible implementation manners.

[0035] In a fourth aspect, the embodiments of the present application provide an electronic device, including a processor and a memory, the memory being configured to store a computer program, and the processor being configured to invoke and run the computer program stored in the memory to perform the method in the first aspect or the possible implementation manners.

[0036] In a fifth aspect, the embodiments of the present application provide a computer readable storage medium, configured to store computer program instructions, and the computer program instructions make a computer perform the method in the first aspect or the possible implementation manners.

[0037] In a sixth aspect, the embodiments of the present application provide a computer program product, including computer program instructions, and the computer program instructions make a computer perform the method in the first aspect or the possible implementation manners.

[0038] In a seventh aspect, the embodiments of the present application provide a device, including a logic circuit and an input and output interface, the input and output interface is configured to receive signals from other communication devices outside the device and transmit the signals to the logic circuit or send signals from the logic circuit to other communication devices outside the device, and the logic circuit is configured to execute code instructions to implement the method in the first aspect or the possible implementation manners.

[0039] In an eighth aspect, the embodiments of the present application provide a vehicle, including a control device, the control device is configured to perform the method in the first aspect or the possible implementation manners of the first aspect.

[0040] In a possible implementation, the vehicle further includes a driving execution device, configured to receive the control instruction sent by the control device, and control the vehicle to drive according to the target strategy according to the control instruction. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A schematic diagram of an application scenario of vehicle control provided by an embodiment of the present application;

[0042] Figure 2 A structural schematic diagram of an electronic device provided by an embodiment of the present application;

[0043] Figure 3 A flow schematic diagram of a control method of a vehicle provided by an embodiment of the present application;

[0044] Figure 4 A schematic diagram of strategy space discretization provided by an embodiment of the present application;

[0045] Figures 5a to 5e A schematic diagram of a decision tree provided by an embodiment of the present application;

[0046] Figures 6a to 6c A flow schematic diagram of driving decision provided by an embodiment of the present application;

[0047] Figures 7a to 7c A schematic block diagram of a model architecture provided by an embodiment of the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

[0049] The execution subject of the technical solutions in the embodiments of the present application can be an electronic device, which can be deployed in a movable device or connected to a movable device through a wired or wireless manner. Of course, the electronic device can also be the movable device itself. The movable device can have any appearance, for example, a smart vehicle, a smart robot, etc. The smart vehicle can be an automatic driving vehicle in which all functions are realized as automatic control, or an assisted driving vehicle in which part of the functions are realized as automatic control to provide driving assistance. In the following, the automatic driving vehicle or the assisted driving vehicle is simply referred to as a vehicle.

[0050] When the vehicle is in an automatic driving mode, the environment around the ego vehicle needs to be perceived to perceive the traffic participants such as pedestrians, vehicles, obstacles, etc. that can exist in the surrounding environment. In the following, for the convenience of description, the pedestrians, vehicles, obstacles, etc. in the environment around the ego vehicle are collectively referred to as target objects. Further, the vehicle predicts the future activity trajectories of the traffic participants, and performs driving planning for the ego vehicle according to the driving navigation of the ego vehicle and the predicted trajectories of the traffic participants, and then realizes driving control of the ego vehicle.

[0051] In order to avoid the modal collapse problem of the algorithm and improve the accuracy of the decision result, the embodiments of the present application introduce a two-dimensional strategy space. Each target point in the strategy space can indicate the driving position of the vehicle in the spatial dimension and the driving speed and / or acceleration of the vehicle in the time dimension. In the process of making driving decisions based on vehicle information and environmental information, the target point that determines the driving strategy is determined through the explicit semantics of the strategy space, the modal collapse problem of the algorithm is avoided, and the accuracy of the driving decision is improved.

[0052] The technical solutions provided by the embodiments of the present application will be exemplarily described below with reference to the accompanying drawings.

[0053] Figure 1 A schematic diagram of an application scenario of vehicle control provided by the embodiments of the present application is shown. As shown in the figure, the vehicle control system 100 can include an information acquisition device 110, a driving planning device 120, and a driving execution device 130. The driving planning device 120 is in communication connection with the information acquisition device 110 and the driving execution device 130 respectively. Figure 1

[0054] The information acquisition device 110 can be used to acquire information of the surrounding environment. For example, the information acquisition device 110 can acquire data collected by image sensors, optical sensors, radars, etc., perceive the surrounding environment, acquire information of target objects, topological information of roads, etc., to realize machine vision. The information acquisition device 110 can also be used to acquire information of the ego vehicle. For example, acquire preset navigation information in the vehicle, the trajectory that the vehicle has traveled, the current speed of the vehicle through a speed sensor, the current steering angle of the vehicle through a torque sensor, the current position of the vehicle through a positioning sensor, etc.

[0055] Further, the information acquisition device 110 sends the acquired information to the driving planning device 120. The driving planning device 120 can make driving planning for the vehicle based on the information acquired by the information acquisition device 110. The driving planning device 120 sends the planning result to the driving execution device 130. The driving execution device 130 controls the vehicle to automatically drive according to the planning result.

[0056] ​The control system 100 of the vehicle can be implemented as a vehicle controller or a component deployed in the vehicle controller.

[0057] In a specific implementation, the driving planning apparatus 120 can include a prediction module 121 and a decision planning module 122. The prediction module 121 can be configured to predict the arrival positions, speeds, and trajectories of the ego vehicle and the target object from a current time to a next time. The decision planning module 122 can determine a driving strategy and a trajectory heuristic of the ego vehicle based on the prediction result of the prediction module 121.

[0058] It should be noted that the vehicle control system 100 can include more or fewer modules than those shown in the figure, which is not limited in the present application. Figure 1 The decision planning module 122 can be divided into a decision module and a planning module.

[0059] The division of each unit in the above apparatus is only a logical functional division, and all or part of them can be integrated into one physical entity, or can be physically separated.

[0060] Figure 2 A structural schematic diagram of an electronic device provided by an embodiment of the present application is shown in the figure. Figure 2 As shown in the figure, the electronic device 200 can include a processor 210 and a memory 220, which communicate with each other through an internal connection path. The memory 220 is configured to store instructions, and the processor 210 is configured to execute the instructions stored in the memory 220.

[0061] Optionally, the memory 220 can include a read-only memory and a random access memory, and provide instructions and data to the processor 210. The memory 220 can be a separate device, or can be integrated in the processor 210.

[0062] In some embodiments, the electronic device 200 can further include an input interface 230. The processor 210 can control the input interface 230 to communicate with other devices or chips, and specifically, can obtain information or data sent by other devices or chips.

[0063] In some embodiments, the electronic device 200 can further include an output interface 240. The processor 210 can control the output interface 240 to communicate with other devices or chips, and specifically, can output information or data to other devices or chips.

[0064] The division of each unit in the above apparatus is only a logical functional division, and all or part of them can be integrated into one physical entity, or can be physically separated.

[0065] It should be understood that the processor in the embodiments of the present application can be an integrated circuit chip with a signal processing capability. In the implementation process, the steps of various embodiments of the method described above can be completed by an integrated logic circuit or an instruction in the form of software in the processor. The processor described above can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The disclosed methods, steps and logic block diagrams in the embodiments of the present application can be implemented or performed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor or the like. The steps of the method disclosed in conjunction with the embodiments of the present application can be directly embodied in a hardware code processor for completion, or a combination of hardware and software modules in the code processor. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register or other mature storage medium in the art. The storage medium is located in the storage, and the processor reads the information in the storage, and combines the hardware to complete the steps of the above method.

[0066] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically EPROM (EEPROM) or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM) and direct memory bus random access memory (DRAM). It should be noted that the memory of the system and method described herein is intended to include but not limited to these and any other suitable types of memory.

[0067] The cache line alignment method provided by the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0068] It should be understood that the following is only for the convenience of understanding and description, and the method provided by the embodiments of the present application is described taking a control device as an execution subject, which can be, for example, the vehicle control system 100 in the above-mentioned Figure 1 or the processor 210 in the above-mentioned Figure 2 . The control device can be implemented as a component in an electronic device, such as a chip, a chip system, or other functional modules capable of invoking and executing programs.

[0069] However, it should be understood that this should not constitute any limitation on the execution subject of the method provided by the present application. As long as the method provided by the embodiments of the present application can be implemented by running a program in which the method provided by the embodiments of the present application is recorded, it can be an execution subject of the method provided by the embodiments of the present application.

[0070] Figure 3 A flowchart of a control method 300 of a vehicle provided by the embodiments of the present application is shown in FIG. 3. As shown in FIG. 3, the method 300 can include at least part of the steps S310 to S340. The steps in the method 300 will be described in detail below. Figure 3

[0071] S310, obtaining vehicle information and environment information, the vehicle information including at least one of driving intention information, driving trajectory information or pose information of the vehicle, and the environment information including topological information of a road and / or information of a target object in the road;

[0072] S320, determining a first goal point from a first strategy space according to the vehicle information and the environment information, the coordinates of the first goal point in the first strategy space including a value of a first space goal point and a value of a first time goal point, the first strategy space including M*N goal points, the M*N goal points being determined by M space goal points and N time goal points, the space goal point indicating a position that the vehicle can reach within a preset time or a preset distance, and the time goal point indicating a driving speed of the vehicle in the corresponding space goal point within the preset time, M and N being positive integers greater than 1;

[0073] S330, determining a target strategy according to the first goal point;

[0074] S340, controlling the vehicle to drive according to the target strategy.

[0075] ​The driving intention information in the vehicle information indicates a driving intention of the vehicle, which can be preset navigation information in the vehicle. The driving trajectory information in the vehicle information indicates a historical trajectory of the vehicle. The pose information in the vehicle information indicates a position and an attitude of the vehicle, and can include position information of the vehicle and attitude information of the vehicle, wherein the attitude information can be an attitude of the vehicle in the road, such as a steering angle of the vehicle relative to a lane line, for example, the control device obtains the current position of the vehicle through a positioning sensor and the current steering angle of the vehicle through a torque sensor.

[0076] The information of the target object in the environment information includes position information, shape and size information, and historical trajectory of the target object, and the target object can be obtained based on machine vision. The topological information in the environment information can be obtained based on machine vision, or determined based on preset navigation information, or determined in combination with machine vision and navigation information.

[0077] The first strategy space can be a space pre-constructed for indicating various possible driving strategies of the vehicle. The first strategy space can include a plurality of discrete target points. The first strategy space can be a two-dimensional space including a spatial dimension and a time dimension, and correspondingly, each target point is a two-dimensional coordinate point, and each target point has data of the time dimension and data of the spatial dimension. The data of the spatial dimension of the target point is used to indicate a position that the vehicle can reach within a preset time or a preset distance, and the data of the time dimension of the target point is used to indicate a driving speed of the vehicle to reach the position indicated by the data of the spatial dimension within a preset time.

[0078] For example, the control device can determine an initial strategy space according to the vehicle information and the environment information, and the initial strategy space can be a continuous two-dimensional space including a spatial (horizontal) dimension and a time (vertical) dimension. The initial strategy space can cover a road level, a lane level, and a sub-lane level in the spatial dimension; and the initial strategy space can cover a vehicle speed and an acceleration in the time dimension. For example, the control device can determine an initial strategy space at a preset period, for example, every 5 seconds. Further, the control device can discretize the initial strategy space to obtain M spatial target points and N time target points, thereby obtaining a first strategy space with M*N target points.

[0079] The period length of the preset period can be the same as or different from the preset time. If the processing complexity is reduced and the processing efficiency is improved, the period length of the preset period can be greater than the preset time, which can be one frame or several frames. The control device can periodically determine the first target point with the preset time as the period length, so as to determine the driving strategy of the vehicle in each period and realize the automatic driving control of the sustainability of the vehicle driving.

[0080] In combination Figure 4 As shown in FIG. 4, the M spatial target points can respectively correspond to lateral actions of the vehicle, for indicating left and right directions of the vehicle, and a straight line formed by the M spatial target points can be parallel to a ground plane, and there can be a same interval between each two spatial target points, or the interval between each two spatial target points can be different, which is not limited in the present application. The M spatial target points are sampled at a first distance from the ego vehicle in the Euclidean space, and a spatial plane formed by the M spatial target points is perpendicular to a road direction, and the spatial plane is at the first distance from the vehicle in a time dimension. The first distance can be determined based on a preset time and a current speed of the vehicle. For example, the control device needs to predict a driving strategy of the vehicle in a next frame, and the M spatial target points indicate predicted reachable positions of the vehicle in different directions within a time length of a frame.

[0081] In some scenarios, the current speed of the vehicle can be small or even zero, and the first distance determined based on the preset time and the current speed of the vehicle can be small. In this case, it is impossible to predict a path of the vehicle to the first distance, or in other words, the prediction lacks the value of automatic driving control. Therefore, the first distance of the spatial plane in which the M spatial target points are located from the vehicle in the time dimension can be a preset distance, which can be understood as a minimum distance for path prediction, that is, the first distance is greater than or equal to the preset distance, and when the first distance is greater than the preset distance, the first distance can be determined based on the preset time and the current speed.

[0082] Still referring to FIG. 4, Figure 4 The N time target points can respectively correspond to longitudinal actions of the vehicle, for indicating speeds of the vehicle in the preset time. Each of the M spatial target points corresponds to the N time target points, in other words, the N time target points are a candidate point set on a sampling line connected with the corresponding spatial target point. For example, Figure 4 The N time target points in FIG. 4 are a candidate point set on a sampling line connected with the spatial target point Gs 4. The time target points can determine a driving speed of the vehicle in the preset time, which can be a constant speed or a non-constant speed, and when the vehicle drives at a non-constant speed in the preset time or the driving speed of the vehicle in the preset time is different from the current speed, the time target points can also determine an acceleration of the vehicle in the preset time.

[0083] The M spatial target points and the N time target points can be combined in a first strategy space to obtain M*N target points, and a coordinate of each target point includes a horizontal coordinate of the spatial target point and a vertical coordinate of the time target point. The target points in the strategy space can determine a driving strategy of the vehicle, such as a position and a speed of the vehicle in the preset time.

[0084] The first target point determined from the first strategy space can determine the position and speed of the vehicle at a preset time. The first space target point meets the driving intention of the vehicle, and the position of the vehicle at the first time target point at the preset time is closest to the driving intention. Of course, the first target point is also affected by the information of the target object, so that the vehicle can avoid collision with the target object during driving.

[0085] In the above S320, the control device can determine a first target point from the first strategy space, which meets the driving intention of the vehicle and can achieve effective avoidance of the target object.

[0086] The first target point can indicate the driving position and speed of the vehicle within a preset time, and when the control device drives from the current position to the position determined by the first target point, the driving trajectory between the two positions (the predicted trajectory in the following) can also be predicted. The first target point and the predicted driving trajectory can form a driving strategy of the vehicle within a preset time (i.e., the target strategy in the above).

[0087] Further, the control device controls the vehicle to drive according to the target strategy, so as to realize automatic driving control of the vehicle.

[0088] In the embodiments of the present application, the control device determines the target point that determines the target strategy in the strategy space with clear semantics based on the vehicle information and the environment information, so as to realize the driving control of the vehicle, avoid the modal collapse problem of the algorithm, and improve the accuracy of the driving decision.

[0089] In some embodiments, the control device can determine a first probability distribution according to the vehicle information and the environment information, the first probability distribution including M*N target points for determining the target strategy. The first probability distribution can refer to Table 1 as follows, which shows the probability of M*N target points formed by M space target points (Gs 0 to Gs M-1) and N time target points (Gt 0 to Gt N-1) for determining the target strategy. For example, the probability of the target point formed by the space target point Gs 0 and the time target point Gt 0 is P 0,0 , the probability of the target point formed by the space target point Gs 1 and the time target point Gt 2 is P 2,1 , and so on. The target point formed by the space target point and the time target point is a target point with the value of the space dimension of the space target point and the value of the time dimension of the time target point as coordinates.

[0090] Table 1

[0091]

[0092] In Table 1, the M spatial target points (Gs 0 to Gs M-1) can respectively indicate the lateral action of the vehicle. For example, the spatial target points Gs 0 and Gs 1 can indicate the vehicle to turn left, and Gs 0 indicates a greater degree of left deviation than Gs 1, Gs 2 to Gs M-3 can indicate the vehicle to keep the current driving direction, Gs M-2 and Gs M-1 can indicate the vehicle to turn right, and Gs M-1 indicates a greater degree of right deviation than Gs M-2.

[0093] In Table 1, the N temporal target points (Gt 0 to Gt N-1) can respectively indicate the longitudinal speed (and / or acceleration) of the vehicle. For example, the temporal target points Gt 0 and Gt 1 can indicate the vehicle to accelerate, and Gt 0 indicates a greater speed (and / or acceleration) than Gt 1, Gt 2 to Gt N-3 can indicate the vehicle to keep constant speed, Gt N-2 and Gt N-1 can indicate the vehicle to decelerate, and Gt N-1 indicates a smaller speed (and / or acceleration) than Gt N-2.

[0094] For example, the availability of the M spatial target points constitutes a probability distribution, when the selection of the spatial target point is determined, the temporal target points associated with the spatial target point form a conditional probability distribution based on the control target point, and the M spatial target points and the N temporal target points jointly constitute a joint probability distribution. The control device determines the probability distribution of the M spatial target points according to the vehicle information and the environment information, and determines the conditional probability distribution of the N temporal target points under each of the M spatial target points according to the probability of each of the M spatial target points, and determines the first probability distribution according to the conditional probability distribution of the N temporal target points under the M spatial target points, for example, the first probability distribution is composed of the conditional probability distribution of the N temporal target points under the M spatial target points.

[0095] Further, the control device determines the first target point from the first strategy space according to the first probability distribution. In some embodiments, the control device can determine the first probability distribution through a neural network model. For example, the vehicle information, the environment information and the first strategy space (such as the information of the M by N target points in the first strategy space) are input into the first model, the first model outputs the first probability distribution, and the first model is trained based on the neural network model. The following will be exemplarily described by using the first model.

[0096] In the first implementation, the first target point can be the target point corresponding to the maximum value in the first probability distribution, such as the maximum value P 1,1 Gs 1 and Gt 1 form the target point D 1,1 as the first target point. It should be understood that the target point corresponding to the maximum value in the first probability distribution represents the global optimal solution of the first target point.

[0097] The control device takes the target point corresponding to the maximum value in the first probability distribution as a first target point, and further determines a target strategy based on the first target point. For example, the control device can determine a lateral action and a longitudinal speed solution of the vehicle according to the target point of the maximum value in the first probability distribution output by the first model. The control device can generate the target strategy by a rolling optimization method, and further control the vehicle to automatically drive.

[0098] In some embodiments, in order to ensure road safety, the control device can determine a predicted trajectory based on the target point corresponding to the maximum value in the first probability distribution, and perform a road safety evaluation on the predicted trajectory. It should be understood that the control device can use any algorithm or model to perform the road safety evaluation, which is not limited in the present application.

[0099] Generally, when performing the road safety evaluation on the predicted trajectory, it is necessary to determine whether there is a collision risk between the vehicle and at least one target object around the vehicle during driving. Therefore, the control device can further predict a driving trajectory of at least one target object (which can include part or all of the target objects around the vehicle) around the vehicle, which can be expressed as a predicted trajectory of the target object, and further perform the road safety evaluation according to the predicted trajectory of the vehicle itself and the predicted trajectory of the at least one target object.

[0100] During the road safety evaluation on the predicted trajectory, the control device can perform the road safety evaluation in combination with an interaction label of the at least one target object. For each target object, the interaction label indicates an interaction strategy of the vehicle to the target object, which includes at least one of yielding, cutting in, and detouring, for example. For example, there is a target object not far in front of the vehicle, and the predicted trajectory of the target object is straight. If the predicted trajectory of the vehicle is straight, the interaction label indicates that detouring is safer than cutting in.

[0101] The control device can determine the predicted trajectory of the vehicle according to vehicle information and environmental information. For example, the control device can output the predicted trajectory of the vehicle through the first model, or the control device can perform trajectory prediction according to the vehicle information and the environmental information for the target point corresponding to the maximum value in the first probability distribution after the first model outputs the first probability distribution.

[0102] In some embodiments, the control device can determine the predicted trajectory of the vehicle in combination with navigation information. As described above, the navigation information can reflect the planned driving path of the vehicle, and therefore, the determination of the predicted trajectory of the vehicle in combination with the navigation information on the basis of the vehicle information and the environmental information can improve the accuracy of trajectory prediction.

[0103] The control device can determine the predicted trajectory of the at least one target object according to the vehicle information and the environment information. Illustratively, the control device can output the predicted trajectory of the at least one target object through the first model. Similarly, the control device can determine the interaction label of the at least one target object according to the vehicle information and the environment information. Illustratively, the control device can output the interaction label of the at least one target object through the first model.

[0104] The result of the driving safety evaluation may, for example, be a numerical value indicating safety, with a higher safety resulting in a higher numerical value of the safety evaluation result, or a lower safety resulting in a lower numerical value of the safety evaluation result, which is not limited in the present application. When the driving safety evaluation passes, the target point corresponding to the maximum value in the first probability distribution is taken as the first target point; when the driving safety evaluation fails, the target strategy determined by other algorithms (such as a decision tree algorithm) is used to control the vehicle.

[0105] In the second implementation, the first target point can be determined from K candidate target points in the first strategy space, K being an integer greater than 1, and the K candidate target points respectively corresponding to K probability values in the first probability distribution. The K probability values include K maximum values, such as P 1,1 is a maximum value 1, P 2,2 is a maximum value 2, and P N-3，M-3 is a maximum value 3. Then the K candidate target points include a target point D 1,1 formed by Gs 1 and Gt 1, a target point D 2,2 formed by Gs 2 and Gt 2, and a target point D M-3,M-3 formed by Gs M-3 and Gt M-3. 1,1 is a maximum value 1, P 2,2 is a maximum value 2, and P N-3，M-3 is a maximum value 3, and the probability values within the preset range include P 1,1 , P 0,1 , P 1,0 , P 2,1 , P 1,2 , P 2,2 , P 1,2 , P 2,1 , P 3,2 , P 2,3 , P N-3,M-3 , P N-4,M-3 , P N-4,M-3 , P N-2,M-3 , P N-3,M-2 Then the K candidate target points include a target point D 1,1, D 2,2 , D M-3,M-3 , D 0,1 , D 1,0 , D 2,1 , D 1,2 , D 2,2 , D 1,2 , D 2,1 , D 3,2 , D 2,3 , D N-3,M-3 , D N-4,M-3 , D N-4,M-3 , D N-2,M-3 , D N-3,M-2 .

[0106] As a first example of the second implementation manner, the control device can determine, based on each of the K candidate target points, a predicted trajectory corresponding to the candidate target point, and perform driving safety evaluation on the predicted trajectories respectively corresponding to the K candidate target points, and take the candidate target point with higher driving safety as the first target point.

[0107] As a second example of the second implementation manner, the control device can determine the first target point from the K candidate target points in combination with a decision tree algorithm. For example, for each of the K candidate target points, the control device determines a path of the candidate target point in the decision tree, and determines a generation value of the candidate target point according to the path of the candidate target point in the decision tree, and takes the candidate target point with the minimum generation value as the first target point. It should be understood that each candidate target point has at least one path in the decision tree. The control device inputs information of nodes on each path in the decision tree into a cost function (or called value function) to obtain a generation value of the path, and takes the candidate target point on the path with the minimum generation value as the first target point.

[0108] As a third example of the second implementation manner, the control device can determine the first target point from the K candidate target points in combination with a joint optimization algorithm. For example, the control device inputs the K candidate target points into the joint optimization algorithm to obtain the first target point. Further, the control device can also input the K candidate target points and decision-related information into the joint optimization algorithm, the decision-related information including at least one of a predicted trajectory of the vehicle corresponding to each candidate target point, a predicted trajectory of at least one target object, and an interaction label of at least one target object, or the decision-related information can include vehicle information and environment information.

[0109] As a third implementation manner, the control device can determine the first target point from the M*N target points in the first strategy space in combination with a decision tree algorithm.

[0110] As follows Figures 5a to 5eThe implementation of the decision tree algorithm shown can be applied to either the second example or the third implementation of the second implementation method described above. In the second implementation method, the control device calculates the cost value for each path in the decision tree that includes K candidate target points, and selects the target point on the path with the smallest cost value as the first target point. In the third implementation method, the control device needs to calculate the cost value for each path in the decision tree, and then selects the target point on the path with the smallest cost value as the first target point. The third implementation method has higher human-likeness, while the second implementation method has higher processing efficiency.

[0111] Figures 5a to 5e This embodiment of the application uses a decision tree algorithm to determine the first target point and constructs a decision tree. Figures 5a to 5e In this model, the constructed decision tree comprises three levels of nodes: the first-level node is the parent node of the second-level node, and the second-level node is the parent node of the third-level node. Specifically, the first-level node represents the probability of a target point in the first policy space being used to determine the target policy; the second-level node represents the predicted trajectory of the vehicle; and the third-level node represents the interaction policy of at least one target object. This interaction policy includes at least one of yielding, cutting off, or detouring, meaning that the interaction policy of a target object instructs the vehicle to yield, cut off, or detour around that target object. The interaction policy of at least one target object can be determined based on vehicle information and environmental information. In some embodiments, the first model can output the interaction policy between the vehicle and at least one target object.

[0112] like Figure 5a As shown, in the decision tree, the first sub-level node of the first level represents the probability of a spatial target point in the first policy space being used to determine the target policy. The second sub-level node of the first level represents the probability of a temporal target point in the first policy space being used to determine the target policy. The second sub-level node of the first level is a temporal target point expanded based on the spatial target point, and the first sub-level node of the first level is the parent node of the second sub-level node of the first level. The nodes of the second level represent the predicted trajectory of the vehicle, and the nodes of the second level represent various possible predicted trajectories for each target point expanded based on the corresponding temporal target point. The first sub-level node of the third level represents the interaction label of at least one target object, and the first sub-level node of the third level is the interaction label expanded based on the predicted trajectory. The second sub-level node of the third level represents the predicted trajectory of at least one target object, and the second sub-level node of the third level is the predicted trajectory of at least one target object expanded based on the interaction label.

[0113] The expansion order of each node when the control device constructs the decision tree and the search order when the decision tree algorithm is adopted can be executed according to the order from high to low of the probability values in the first probability distribution. The higher the probability value in the first probability distribution is, the smaller the generation value of the path related to the corresponding target point is, so that the branch corresponding to the target point with a large probability value is more likely to be hit.

[0114] The first model can output at least one of the predicted trajectory, the interaction label of the at least one target object, or the predicted trajectory of the at least one target object; or the first model can perform refined sampling on at least one of the predicted trajectory, the interaction label of the at least one target object, or the predicted trajectory of the at least one target object based on an output result of the first model; or the first model can perform joint optimization on at least one of the predicted trajectory, the interaction label of the at least one target object, or the predicted trajectory of the at least one target object based on the output of the first model. The present application does not limit this.

[0115] Figure 5b The decision tree shown in Figure 5a is different from the decision tree shown in in that when there are multiple target objects, the information of the multiple target objects is not expanded within one level, but is expanded layer by layer after being sorted based on a sorting rule, for example, is expanded in the order from close to far according to the distance between the target objects and the vehicle. The nodes of one sub-level of the third level indicate the information of one target object. Specifically, the nodes of the first sub-level of the third level represent the interaction label of a first target object in the at least one target object, the second sub-level of the third level represents the predicted trajectory of the first target object, the nodes of the third sub-level of the third level represent the interaction label of a second target object in the at least one target object, and the fourth sub-level of the third level represents the predicted trajectory of the second target object, and so on. The number of sub-levels of the third level can be 2 times the number of target objects.

[0116] Figure 5a On the basis of Figure 5b , the first level of the decision tree can be replaced by the first level shown in Figure 5c , Figure 5c On the basis of replacing the first level with Figure 5a , it should be understood that Figure 5c the first level shown in Figure 5b can also replace the nodes of the first level in Figure 5c . As shown in Figure 5a , the nodes of the first level represent the probabilities of the target points of the first strategy space for determining the target strategy. The two sub-levels of the first level shown in are combined, and the nodes of the target points are directly tiled and expanded, and then the nodes of the predicted trajectory of the vehicle are expanded based on the target points.

[0117] Figure 5d WithFigure 5b The difference lies in the fact that the third level of the decision tree does not include the sub-levels of interactive labels indicating target objects. The nodes of the third level, which indicate the predicted trajectory of the target object, are expanded based on the predicted trajectory of the vehicle, and a predicted trajectory of one target object is expanded in each sub-level of the third level. Target objects can be sorted according to their ordered trajectories, and the predicted trajectory of each target object can be expanded layer by layer, for example, based on the distance between the target object and the vehicle, expanding the sub-levels of the third level in the order of proximity.

[0118] Figure 5e and Figure 5a The difference lies in the fact that the third level of the decision tree does not include a sub-level that indicates the interactive labels of the target object. The nodes of the third level, which indicate the predicted trajectory of at least one target object, are expanded based on the predicted trajectory of the vehicle.

[0119] Figure 5d and Figure 5e The first level shown can also be replaced with Figure 5c The first level in the system.

[0120] In this embodiment of the application, for Figures 5a to 5e The number of nodes at each level in the decision tree shown is not limited.

[0121] As a fourth implementation method, the control device can combine a joint optimization algorithm to determine the first target point from M multiplied by N target points in the first strategy space.

[0122] The technical features in the first and second implementation methods described above can be combined to achieve the driving decision determination scheme in the embodiments of this application.

[0123] Using white-box algorithms such as decision trees for driving decisions has several drawbacks. First, it can lead to poor human-likeness in driving strategies. Second, the number of searches and computational demands increase exponentially with the dimensionality of the problem definition, resulting in low processing efficiency. Third, the decision-making process involves a large number of manual adjustments to parameters based on experience, leading to poor usability. Therefore, this application proposes using a neural network model for driving decisions. This approach improves the human-likeness of the decisions and enables rapid decision-making even with a high degree of problem dimensionality, thus increasing processing efficiency. Furthermore, by replacing manual parameter tuning with model training, it solves the problems of high complexity in cost function parameter definition and large workload in parameter tuning. Finally, based on the clear semantics of the first policy space in this application, it addresses the problem of modal collapse during the neural network model prediction process.

[0124] As previously mentioned, during the driving decision-making process, the control device can input vehicle information, environmental information, and information from M multiplied by N target points into the first model. The first model can be a multi-level or multi-task network, and the control device outputs at least one of the following through the first model:

[0125] First probability distribution;

[0126] The predicted trajectories corresponding to M multiplied by N target points;

[0127] At least one interactive label for the target object;

[0128] The predicted trajectory of at least one target object.

[0129] For example, in combination Figure 6a As shown, the prediction module (such as Figure 1 The prediction module 121) may include a spatiotemporal strategy discretization unit and a first model. The spatiotemporal strategy discretization unit acquires an initial strategy space, which can be a continuous strategy space determined based on vehicle information and environmental information. The spatiotemporal strategy discretization unit can discretize the initial measurement space in both the spatial (lateral) dimension and the temporal (vertical) dimension, respectively, obtaining M spatial target points through spatial dimension discretization and N temporal target points through temporal dimension discretization.

[0130] Furthermore, the spatiotemporal strategy discrete unit takes M spatial target points and N temporal target points, along with acquired vehicle information and environmental information, as inputs to the first model. The first model predicts the spatial strategy probability distribution based on the M spatial target points, vehicle information, and environmental information; that is, the probability of each of the M spatial target points being used to determine the target strategy. The first model also predicts the spatiotemporal strategy probability distribution based on the N temporal target points and the spatial strategy probability distribution, which is the first probability distribution mentioned earlier. Finally, the first model predicts the vehicle's trajectory and the trajectory of at least one target object based on the first probability distribution.

[0131] Furthermore, the first probability distribution output by the first model, the predicted trajectory of the vehicle, and the predicted trajectory of at least one target object are input into the decision planning module. The decision planning module can employ white-box algorithms (such as decision tree algorithms, joint optimization algorithms, etc.) to make decisions and plans based on the data output by the first model.

[0132] For example, the decision planning module performs a strategy space search to determine an optimal solution (e.g., a maximum value in the first probability distribution) of the spatiotemporal strategy (i.e., a strategy corresponding to the target point), performs a driving safety evaluation on the spatiotemporal strategy, and performs driving control based on the spatiotemporal strategy if the evaluation passes. If the evaluation fails, the decision planning module continues to perform a strategy space search to determine at least one maximum value of the spatiotemporal strategy, performs a safety evaluation on the spatiotemporal strategy corresponding to the at least one maximum value, and selects the spatiotemporal strategy that passes the evaluation or has the highest evaluation score as the spatiotemporal strategy for driving control. If the evaluation still fails, the decision planning module continues to perform a strategy space search to determine a neighborhood of the maximum value of the spatiotemporal strategy, and then performs a driving safety evaluation on each spatiotemporal strategy corresponding to a probability value in the neighborhood, selects the spatiotemporal strategy that passes the evaluation or has the highest evaluation score, and uses a white-box algorithm to determine a spatiotemporal strategy based on the vehicle information and the environmental information to implement a fallback decision.

[0133] For example, the decision planning module performs trajectory planning based on the output of the first model. The predicted trajectory output by the first model can be used as an initial solution to accelerate the joint optimization algorithm, and the predicted trajectory of at least one target object output by the first model can be used as a boundary constraint to ensure driving safety. The decision planning module performs a driving safety evaluation on the predicted trajectory, and uses the predicted trajectory for driving control if the evaluation passes. If the evaluation fails, the decision planning module uses a white-box algorithm to predict a trajectory based on the vehicle information and the environmental information to implement a fallback planning.

[0134] Figure 6b Another implementation is shown, Figure 6b The network level in the first model shown is different from Figure 6a The network level in the first model shown is different from Figure 6b In the first model shown, the level of predicting a spatiotemporal strategy probability distribution and the level of predicting a temporal strategy probability distribution are combined into a level of predicting a spatiotemporal strategy distribution. The first model is input with M spatial target points and N temporal target points, vehicle information, and environmental information, determines a matrix of spatiotemporal strategy probability distribution, and then predicts a driving trajectory of the vehicle and a driving trajectory of at least one target object based on the matrix of spatiotemporal strategy probability distribution.

[0135] Figure 6c Another implementation is shown, Figure 6c The difference between the first model shown and Figure 6b is that Figure 6cThe middle spatiotemporal strategy discrete unit discretizes the initial strategy space into a spatiotemporal target point matrix, which is composed of features in M spatial dimensions and features in N temporal dimensions. Then, the spatiotemporal target point matrix, vehicle information and environment information are input into the first model, and the first model can determine a matrix of spatiotemporal strategy probability distribution, and then predict the driving trajectory of the vehicle and the driving trajectory of at least one target object according to the matrix of spatiotemporal strategy probability distribution.

[0136] In order to enable the first model to achieve accurate prediction, an embodiment of the present application provides a model training scheme. In the process of model training, the control device can obtain a plurality of training samples, the training samples including input data and expected output data, the expected output data being used to indicate the expected output of the neural network model when inputting the input data. The expected output data includes at least one of a spatial target point true value, a temporal target point true value, a target point true value, a trajectory true value of the vehicle, a trajectory true value of the target object and an interaction label true value.

[0137] The input data includes M'N' target points of the second strategy space, vehicle information and environment information in the human driving interaction process, and M' and N' are both positive integers greater than 1. The second strategy space can be determined based on the vehicle information and the environment information in the input data, for example, the control device determines the initial strategy space according to the vehicle information and the environment information, and discretizes the initial strategy space to obtain the second strategy space. M' can be the same as M or different from M, and N' can be the same as N or different from N. The vehicle information and the environment information in the input data can be obtained in the non-automatic driving process, i.e., the vehicle information and the environment information obtained when the driver controls the vehicle to drive.

[0138] The control device inputs the plurality of training samples into the neural network model, trains the neural network model, and obtains the first model after the loss value converges to a preset value, the loss value being determined based on the expected output data and a predicted value of the input data by the neural network model.

[0139] The various training stages of the model training process in the embodiment of the present application will be exemplarily described below with reference to the accompanying drawings.

[0140] Referring to Figure 7a The encoder in the neural network model can perform feature vector conversion on the input data to model the interaction between the vehicle and the road, and the interaction between the vehicle and the target object. The decoder in the neural network model performs prediction (such as probability prediction, trajectory prediction, and interaction label prediction) based on the feature vector, and returns the classification loss value according to the true value, so as to train the encoder and the decoder in the neural network model in an iterative manner to obtain the first model.

[0141] AsFigure 7a As shown, in the embodiments of the present application, the model training process can include a first training phase, a second training phase and a fourth training phase.

[0142] In the first training phase, the control device predicts the probabilities corresponding to the M' spatial target points through the neural network model, and determines a first loss value according to the spatial target point with the highest probability in the M' spatial target points and the spatial target point true value. Through iterative training, the first training phase ends when the first loss value converges to a first preset value, and the second training phase is entered.

[0143] In the second training phase, the control device predicts the probabilities corresponding to the N' time target points under the spatial target point true value through the neural network model, which can be the neural network model trained in the first training phase, and determines a second loss value according to the time target point with the highest probability in the N' time target points and the time target point true value. Through iterative training, the second training phase ends when the second loss value converges to a second preset value, and the fourth training phase is entered.

[0144] In the fourth training phase, the control device predicts the trajectory corresponding to the time target point true value through the neural network model, which can be the neural network model trained in the first training phase and the second training phase, and determines a fourth loss value according to the trajectory corresponding to the time target point true value and the trajectory true value of the vehicle. Through iterative training, the fourth training phase ends when the fourth loss value converges to a fourth preset value, and the training process ends to obtain the first model.

[0145] In some embodiments, the fourth training phase is an optional training phase. For example, the control device can end the model training process after completing the first training phase and the second training phase to obtain the first model. In this case, the first model is used to predict the first probability distribution.

[0146] Figure 7b As shown, another model architecture is shown, Figure 7b The difference between Figure 7a is that the control device simultaneously predicts the time strategy and the space strategy in one training phase. In Figure 7b As shown in the example, the model training process can include a third training phase and a fourth training phase.

[0147] In the third training phase, the control device predicts the joint probabilities corresponding to the M' spatial target points and the N' time target points through the neural network model, and determines a third loss value according to the target point with the highest joint probability in the M'N' target points and the target point true value. Through iterative training, the third training phase ends when the third loss value converges to a third preset value.

[0148] In the fourth training stage, the control device predicts a trajectory corresponding to the ground truth of the target point through a neural network model, which can be the neural network model trained in the first and second training stages, and determines a fourth loss value according to the trajectory corresponding to the ground truth of the target point and the ground truth of the trajectory. The fourth training stage ends when the fourth loss value converges to a fourth preset value through iterative training.

[0149] Figure 7c Another model architecture is shown as follows, Figure 7c The model architecture shown above is different from Figure 7a The model architecture shown above has a different network cascade structure. Figure 7a In the first model architecture, the neural network model first decodes the probability of the time target point based on the probability prediction of the space target point and the feature vector, and then decodes the trajectory of the vehicle, while Figure 7c In the second model architecture, the neural network model simultaneously decodes the probability of the time target point and the trajectory of the vehicle based on the space target point prediction and the feature vector.

[0150] As shown above, Figure 7c The model training process can include a first training stage, a second training stage, and a fourth training stage. The first training stage and the second training stage are consistent with the embodiments shown above, and will not be described again here. In the fourth training stage in the embodiments shown above, Figure 7a The control device can predict a trajectory corresponding to the ground truth of the space target point through a neural network model after completing the first training stage of the training process, which can be the neural network model trained in the first training stage, and determine a fourth loss value according to the trajectory corresponding to the ground truth of the space target point and the ground truth of the trajectory. The fourth training stage ends when the fourth loss value converges to a fourth preset value through iterative training. Figure 7c

[0151] Figure 7b In the model training manner in the embodiments shown above, the efficiency of model training is higher, and the first model trained has higher processing efficiency when implementing prediction, Figure 7c In the model training manner in the embodiments shown above, the accuracy of the first model trained is higher. Figure 7a

[0152] The model training process in the embodiments of the present application can further include a fifth training stage. In the fifth training stage, the control device predicts a trajectory of the target object through a neural network model, and determines a fifth loss value according to the trajectory of the target object and the ground truth of the trajectory of the target object. The fifth training stage ends when the fifth loss value converges to a fifth preset value through iterative training. The fifth training stage can be executed after any of the first to fourth training stages, and the neural network model trained in the previous training stage is used to continue the model training process.​​

[0153] The model training process in the embodiments of the present application can further include a sixth training stage. In the sixth training stage, the control device predicts the interaction label of the target object through the neural network model, determines a sixth loss value according to the interaction label of the target object and the interaction label true value, and ends the sixth training stage when the sixth loss value converges to a sixth preset value through iterative training. The sixth training stage can be executed after any of the first training stage to the fourth training stage is completed, and the neural network model trained in the previous training stage is used to continue the model training process.

[0154] The fifth training stage and the sixth training stage can be executed simultaneously or in sequence, and the order of the fifth training stage and the sixth training stage is not limited when they are executed in sequence.

[0155] The embodiments of the present application further provide a computer readable storage medium for storing a computer program.

[0156] In some embodiments, the computer readable storage medium can be applied to the processing device in the embodiments of the present application, and the computer program makes the computer execute the corresponding procedures in the various methods of the embodiments of the present application. For brevity, details are not repeated here.

[0157] The embodiments of the present application further provide a computer program product including computer program instructions.

[0158] In some embodiments, the computer program product can be applied to the processing device in the embodiments of the present application, and the computer program instructions make the computer execute the corresponding procedures in the various methods of the embodiments of the present application. For brevity, details are not repeated here.

[0159] The embodiments of the present application further provide a computer program.

[0160] In some embodiments, the computer program can be applied to the processing device in the embodiments of the present application, and when the computer program runs on the computer, it makes the computer execute the corresponding procedures in the various methods of the embodiments of the present application. For brevity, details are not repeated here.

[0161] The embodiments of the present application further provide an intelligent vehicle.

[0162] In some embodiments, the vehicle includes the control device in the embodiments of the present application. The control device is used to execute the corresponding procedures of the embodiments of the present application.

[0163] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0164] The above description is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A control method of a vehicle, characterized by, The method comprises: obtaining vehicle information and environment information, the vehicle information comprising at least one of driving intention information, driving trajectory information or pose information of the vehicle, and the environment information comprising topological information of a road and / or information of a target object in the road; determining a first target point from a first strategy space according to the vehicle information and the environment information, the coordinates of the first target point in the first strategy space comprising a value of a first space target point and a value of a first time target point, the first strategy space comprising M*N target points determined by M space target points and N time target points, the space target point indicating a position reachable by the vehicle within a preset time or a preset distance, and the time target point indicating a driving speed and / or acceleration of the vehicle within the preset time in a direction of the corresponding space target point, M and N being positive integers greater than 1; determining a target strategy according to the first target point; and controlling the vehicle to drive according to the target strategy.

2. The method of claim 1, wherein, The space surface formed by the M space target points in the Euclidean space is perpendicular to the road direction, and the space surface is away from the vehicle by a first distance, the first distance being greater than or equal to the preset distance.

3. The method of claim 2, wherein, The first distance is determined based on the preset time and a current speed of the vehicle.

4. The method according to any one of claims 1 to 3, characterized in that, The determining of the first target point from the first strategy space according to the vehicle information and the environment information comprises: determining a first probability distribution according to the vehicle information and the environment information, the first probability distribution comprising probabilities of the M*N target points for determining the target strategy; and determining the first target point from the first strategy space according to the first probability distribution.

5. The method of claim 4, wherein, The determining of the first target point from the first strategy space according to the first probability distribution comprises: taking a target point corresponding to a maximum value in the first probability distribution as the first target point; or determining K candidate target points from the first strategy space according to the first probability distribution, and determining the first target point from the K candidate target points, K being an integer greater than 1; wherein the K candidate target points respectively correspond to K probability values in the first probability distribution, the K probability values comprising a maximum value, or the K probability values comprising the maximum value and probability values within a preset range of the maximum value, or the K probability values being the first K probability values in the first probability distribution in descending order; or adopting a decision tree algorithm to determine the first target point from the first strategy space according to the first probability distribution; or adopting a joint optimization algorithm to determine the first target point from the first strategy space according to the first probability distribution.

6. The method of claim 5, wherein, The taking of the target point corresponding to the maximum value in the first probability distribution as the first target point comprises: determining a predicted trajectory based on the target point corresponding to the maximum value in the first probability distribution; performing driving safety evaluation on the predicted trajectory; when the driving safety evaluation passes, taking the target point corresponding to the maximum value in the first probability distribution as the first target point.

7. The method of claim 5, wherein, The determining the first target point from the at least two candidate target points comprises: performing driving safety evaluation on the predicted trajectories respectively corresponding to the K candidate target points, and taking a candidate target point with higher driving safety as the first target point; or adopting a decision tree algorithm to determine the first target point from the K candidate target points; or adopting a joint optimization algorithm to determine the first target point from the K candidate target points.

8. The method of claim 7, wherein, The adopting the decision tree algorithm to determine the first target point from the K candidate target points comprises: determining, for each of the K candidate target points, a path of the candidate target point in the decision tree; determining a generation value of the candidate target point according to the path of the candidate target point in the decision tree; taking a candidate target point with the minimum generation value as the first target point.

9. The method according to any one of claims 5 to 8, characterized in that, The decision tree comprises nodes of a first level, nodes of a second level, and nodes of a third level, the nodes of the first level being parent nodes of the nodes of the second level, and the nodes of the second level being parent nodes of the nodes of the third level; the nodes of the first level represent probabilities of target points of the first strategy space for determining the target strategy; the nodes of the second level represent predicted trajectories of the vehicle; the nodes of the third level represent interaction labels and / or predicted trajectories of at least one target object, the interaction labels being used to indicate interaction strategies of the vehicle to the target object, the interaction strategies comprising at least one of yielding, taking, and bypassing.

10. The method of claim 9, wherein, nodes of a first sub-level of the first level represent probabilities of spatial target points of the first strategy space for determining the target strategy, nodes of a second sub-level of the first level represent probabilities of temporal target points of the first strategy space for determining the target strategy, and the nodes of the first sub-level of the first level are parent nodes of the nodes of the second sub-level of the first level.

11. The method of claim 9, wherein the nodes of a first sub-level of the third level represent interaction labels of the at least one target object, and the nodes of a second sub-level of the third level represent predicted trajectories of the at least one target object; or the nodes of a first sub-level of the third level represent interaction labels of a first target object of the at least one target object, the second sub-level of the third level represents a predicted trajectory of the first target object, the nodes of a third sub-level of the third level represent interaction labels of a second target object of the at least one target object, and the fourth sub-level of the third level represents a predicted trajectory of the second target object, and so on; or the nodes of a first sub-level of the third level represent predicted trajectories of a first target object of the at least one target object, and the nodes of a second sub-level of the third level represent predicted trajectories of a second target object of the at least one target object, and so on. The method further comprises:

12. The method according to any of claims 6-8, 10-11, characterized by, ​ According to the vehicle information and the environment information, a predicted trajectory and an interaction label of at least one target object are determined, the interaction label being used to indicate an interaction strategy of the vehicle to the target object, the interaction strategy including at least one of yielding, taking the road, and bypassing, and the predicted trajectory and the interaction label of the target object are used to perform driving safety evaluation.

13. The method of claim 12, wherein, The determining, according to the vehicle information and the environment information, of the predicted trajectory and the interaction label of the at least one target object includes: According to the vehicle information, the environment information, and navigation information, a predicted trajectory and an interaction label of at least one target object are determined, the navigation information indicating a planned driving path of the vehicle.

14. The method of claim 4, wherein, The determining, according to the vehicle information and the environment information, of the first probability distribution includes: According to the vehicle information and the environment information, a probability distribution of the M spatial target points is determined. For a probability of each of the M spatial target points, a conditional probability distribution of the N temporal target points at the spatial target point is determined. According to the conditional probability distribution of the N temporal target points at the M spatial target points, the first probability distribution is determined.

15. The method of any one of claims 1-3, 5-8, 10-11, wherein, Further comprising: According to the vehicle information and the environment information, an initial strategy space is determined. The initial strategy space is discretized to obtain the M spatial target points and the N temporal target points.

16. The method of any one of claims 1-3, 5-8, 10-11, wherein, The determining, according to the vehicle information and the environment information, of the first target point from the first strategy space includes: The vehicle information, the environment information, and the M by N target points are input into a first model, and at least one of the following is output by the first model: a first probability distribution; a predicted trajectory corresponding to each of the M by N target points; an interaction label of at least one target object; a predicted trajectory of at least one target object; The first model is obtained by training a neural network model.

17. The method of claim 16, wherein, Further comprising: A plurality of training samples are obtained, the training samples including input data and expected output data, the expected output data being used to indicate an expected output of the neural network model when the input data is input, the input data including M' by N' target points of a second strategy space, vehicle information, and environment information in a human driving interaction process, the expected output data including at least one of spatial target point true values, temporal target point true values, target point true values, a trajectory true value of the vehicle, a trajectory true value of the target object, and an interaction label true value, M' and N' being positive integers greater than 1; The plurality of training samples are input into the neural network model, the neural network model is trained, and after a loss value converges to a preset value, the first model is obtained, the loss value being determined based on the expected output data and a predicted value of the input data by the neural network model.

18. The method of claim 17, wherein, The inputting of the plurality of training samples into the neural network model and the training of the neural network model, and after the loss value converges to the preset value, the obtaining of the first model include: The plurality of training samples are input into the neural network model, and the first model is obtained through the following training process. The training process comprises a first training stage and a second training stage; In the first training stage, the probabilities corresponding to the M' spatial target points are predicted by the neural network model, and a first loss value is determined according to the spatial target point with the highest probability in the M' spatial target points and the spatial target point true value. The first training stage ends when the first loss value converges to a first preset value through iterative training. In the second training stage, the probabilities corresponding to the N' time target points under the spatial target point true value are predicted by the neural network model, and a second loss value is determined according to the time target point with the highest probability in the N' time target points and the time target point true value. The second training stage ends when the second loss value converges to a second preset value through iterative training. Or, The training process comprises a third training stage; In the third training stage, the joint probabilities corresponding to the M' spatial target points and the N' time target points are predicted by the neural network model, and a third loss value is determined according to the target point with the highest joint probability in the M'N' target points and the target point true value. The third training stage ends when the third loss value converges to a third preset value through iterative training.

19. The method of claim 18, wherein, The training process further comprises a fourth training stage: After the first training stage of the training process ends, the trajectory corresponding to the spatial target point true value is predicted by the neural network model; or, After the second training stage of the training process ends, the trajectory corresponding to the time target point true value is predicted by the neural network model; or, After the third training stage of the training process ends, the trajectory corresponding to the target point true value is predicted by the neural network model; A fourth loss value is determined according to the trajectory and the trajectory true value. The fourth training stage ends when the fourth loss value converges to a fourth preset value through iterative training.

20. The method of claim 18 or 19, wherein, The training process further comprises a fifth training stage and / or a sixth training stage; In the fifth training stage, the trajectory of the target object is predicted by the neural network model, and a fifth loss value is determined according to the trajectory of the target object and the trajectory true value of the target object. The fifth training stage ends when the fifth loss value converges to a fifth preset value through iterative training. In the sixth training stage, the interaction label of the target object is predicted by the neural network model, and a sixth loss value is determined according to the interaction label of the target object and the interaction label true value. The sixth training stage ends when the sixth loss value converges to a sixth preset value through iterative training.

21. A control device characterized by comprising: A module for executing the method of any one of claims 1 to 20.

22. A chip, characterized by A chip comprising: A processor for calling and running computer instructions from a memory, so that the device installed with the chip executes the method of any one of claims 1 to 20.

23. An electronic device, comprising: A chip comprising: A processor and a memory for storing a computer program, the processor being configured to invoke and run the computer program stored in the memory to perform the method of any one of claims 1 to 20.

24. A vehicle characterized by comprising: Comprising: A control device configured to perform the method of any one of claims 1 to 20.

25. The vehicle of claim 24, wherein, Further comprising: a driving execution device configured to: Receive the control instruction sent by the control device; Control the vehicle to drive according to the target strategy according to the control instruction.

26. A computer program product, characterised in that, Computer program instructions for causing a computer to perform the method of any one of claims 1 to 20.

27. A computer-readable storage medium, characterized in that, A computer program for causing a computer to perform the method of any one of claims 1 to 20. A computer program for causing a computer to perform the method of any one of claims 1 to 20.

Images