Vehicle path planning method and device

Abstract

A vehicle path planning method and device, the method comprising: obtaining image information and related data information of traffic participants and road conditions around a test vehicle through a camera, a laser radar, etc.; determining target traffic participants related to the vehicle path planning task using the image information and data information; extracting effective feature vectors of the target traffic participants using a feature extraction model, mapping the feature vectors to a path point coordinate sequence of the vehicle in the future based on a multi-head attention mechanism and reinforcement learning, and planning a path for the vehicle based on the path point coordinate sequence. The present invention can determine target traffic participants related to the path planning task in a complex traffic environment, ignoring other irrelevant traffic participants and their behaviors, and solving the problem that the path planning method in the prior art has weak flexibility and the path planning is easily affected by changes in the surrounding environment. The attention mechanism and reinforcement learning used in the present invention can enhance the interpretability of the model.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a vehicle path planning method and device. Background Art

[0002] In the field of autonomous driving, vehicles obtain information about the surrounding environment through environmental perception modules, and make behavioral decisions, path planning and motion control based on the vehicle's own location information and motion status information, and generate corresponding control instructions; then the control instructions are sent to the vehicle's chassis actuators to achieve autonomous driving of the vehicle. Existing autonomous driving path planning models can be divided into two types: rule-based and learning-based. The rule-based model has a simple structure, strong interpretability, and good real-time performance, but it has weak flexibility and poor robustness when facing complex environments, and the control actions will be stiff due to rule switching, resulting in poor passenger comfort.

[0003] Learning-based path planning methods usually make decisions based on the vehicle's own status information and the status information and road conditions of other traffic participants within a certain range around it. They fail to fully consider: the path planning task of the ego vehicle, i.e. the test vehicle, is not only related to traffic participants that are close to the test vehicle, but is also more related to those traffic participants that are far away from the ego vehicle but conflict with the ego vehicle's future path; not all traffic participants that are close to the ego vehicle are related to the path planning task; traffic participants that do not conflict with the ego vehicle's future path do not need to be taken into consideration; the flexibility of the planning method in the face of complex and sudden environmental changes; and the interpretability of the path planning method. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the existing methods, the present invention provides a vehicle path planning method and device, which can capture specific traffic participants and road conditions that are strongly related to the vehicle path planning task, and optimize the existing vehicle path planning method.

[0005] A first aspect of the present invention provides a vehicle path planning method, comprising:

[0006] Obtain image information and related data information about traffic participants and road conditions around the test vehicle through cameras, laser radar, etc.

[0007] Determining a target traffic participant associated with a vehicle path planning task using the image information and the data information;

[0008] The effective feature vector of the target traffic participant is extracted by using a feature extraction model, and the feature vector is mapped into a path point coordinate sequence of the vehicle in the future based on a multi-head attention mechanism and reinforcement learning, and a path is planned for the vehicle based on the path point coordinate sequence.

[0009] In an optional implementation manner, the determining a target traffic participant related to the vehicle path planning task using the image information and the data information includes:

[0010] The future driving trajectory of the traffic participant is predicted, and the traffic participant whose future driving trajectory conflicts with the future planned path of the vehicle is taken as the target traffic participant.

[0011] In an optional implementation, the extracting the effective feature vector of the target traffic participant by using a feature extraction model includes:

[0012] Using a convolutional neural network model and a multi-layer perceptron to extract effective information from the image information and the data information respectively, the convolutional neural network model outputs a feature vector extracted from the image information as a first feature vector, and the multi-layer perceptron outputs a feature vector extracted from the data information as a second feature vector;

[0013] The first feature vector and the second feature vector are concatenated into a valid feature vector of the target traffic participant.

[0014] In an optional implementation manner, the step of concatenating the first feature vector and the second feature vector into a valid feature vector of the target traffic participant includes:

[0015] Aligning the tensor dimension of the first feature vector with the tensor dimension of the second feature vector; wherein the first feature vector represents image information from a camera, and the second feature vector represents data information acquired by a sensor;

[0016] The first feature vector is regarded as the n+2th second feature vector, the second feature vector is input into the multi-head attention network, and the first feature vector and the second feature vector are concatenated and aggregated into a representation tensor of the target traffic participant and the surrounding road conditions for subsequent input into the multi-head attention mechanism network.

[0017] In an optional embodiment, the multi-head attention model is combined with the Actor-Critic algorithm architecture in reinforcement learning, allowing the input tensor to be variable in dimension and arrangement order; the number of target traffic participants around the test vehicle is variable and the arrangement order of the representation vectors of each target traffic participant is variable; and the feature vector can be mapped into a sequence of path point coordinates of the vehicle in the future.

[0018] In an optional implementation manner, the determining a target traffic participant related to the vehicle path planning task using the image information and the data information includes:

[0019] Obtain image information and data information of traffic participants around the vehicle at the current moment and several previous moments; use the image information and data information of traffic participants around the vehicle at the current moment and several previous moments as input data of the feature extraction model.

[0020] A second aspect of the present invention provides a vehicle path planning device, comprising:

[0021] Acquisition and recognition module, used to obtain image information and related data information of traffic participants and road conditions around the test vehicle through cameras, laser radars, etc.;

[0022] A target determination module, used for determining a target traffic participant related to a vehicle path planning task by using the image information and the data information;

[0023] The path planning module is used to extract the effective feature vector of the target traffic participant by using the feature extraction model, map the feature vector into a path point coordinate sequence of the vehicle in the future based on the multi-head attention mechanism and reinforcement learning, and plan a path for the vehicle based on the path point coordinate sequence.

[0024] The third aspect of the present invention provides a vehicle control method, comprising: obtaining a vehicle planned path determined by the vehicle path planning method described in the second aspect of the present invention, and controlling the vehicle driving according to the vehicle planned path.

[0025] A fourth aspect of the present invention provides an electronic device, comprising:

[0026] At least one processor; and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method described in the first aspect of the embodiment of the present invention.

[0027] According to a fifth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a computer, the method according to the first aspect of the embodiment of the present invention is executed.

[0028] The present invention is a learning-based autonomous driving path planning model that can determine target traffic participants related to the path planning task in a complex and changeable traffic environment, and ignore traffic participants and their behaviors that are not related to the path planning task. The present invention can enhance the decision-making ability of vehicles in complex and changeable traffic environments, especially in the face of emergencies, making path planning more generalizable and flexible; it solves the problems of weak flexibility of path planning tasks and susceptibility of path planning to changes in the surrounding environment in existing technical solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The figure is a flow chart of a vehicle path planning method in an embodiment of the present invention.

[0030] Figure 2 Schematic diagram of the flow of the target traffic participant identification method in an embodiment of the present invention.

[0031] Figure 3 Schematic diagram of a module of a vehicle path planning device in an embodiment of the present invention.

[0032] Figure 4 Schematic diagram of the process of the vehicle control method in the embodiment of the present invention.

[0033] Figure 5 FIG. 1 is a schematic structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0035] It should be understood that the terms "first", "second", and "third" in the claims, specifications, and drawings of the present disclosure are used to distinguish different objects rather than to describe a specific order. The terms "include" and "comprise" used in the specifications and claims of the present disclosure indicate the presence of the described features, entities, steps, operations, but do not exclude the presence or addition of one or more other features, entities, steps, operations, elements, components, and / or their collections.

[0036] In the existing technology, reinforcement learning has been used more to train the motion controller of autonomous vehicles. The common method for path planning is rule-based. However, the surrounding environment in the real world changes rapidly and it is difficult to predict accurately. Once the surrounding environment, such as the behavior of obstacles in front, changes, the decision of the vehicle must change immediately. The learning-based path planning method has stronger generalization ability and flexibility than the rule-based method.

[0037] The input of path planning includes high-precision maps, the status of traffic lights, destination coordinates, the vehicle's own posture, position coordinates and other information. The output of path planning is a function that can map different moments into the corresponding position coordinates of the vehicle, that is, the position that the vehicle needs to reach at a given moment in the future. Therefore, path planning is to calculate a safe and comfortable trajectory for the autonomous vehicle to complete the predetermined driving task. The vehicle maintains an appropriate distance from obstacles during driving to avoid collisions; provides passengers with a comfortable riding experience, such as avoiding excessive acceleration and deceleration, slowing down appropriately on curves to avoid excessive centripetal acceleration, etc.; finally, completing the driving task means that the planned path must reach the destination while complying with traffic rules, not violating laws and regulations, and cannot cause unacceptable driving time due to too low a speed.

[0038] Path planning methods based on deep reinforcement learning usually make decisions based on the state information of the vehicle itself, the state information of other traffic participants within a certain range around it, and the road conditions. They fail to fully consider that the path planning task goal of the vehicle is only related to specific traffic participants, but instead include all traffic participants within a certain range around the vehicle in the model. In addition, the number of traffic participants within a certain range around the vehicle is constantly changing, so the input of the path planning model needs to meet the following two requirements: one is that the dimension is variable, and the other is that the sorting order of the representation vector input network of the traffic participants is variable. At present, the commonly used path planning networks in the field of autonomous driving include convolutional neural networks, multi-layer perceptrons, etc., and these networks cannot meet the above two requirements at the same time.

[0039] The path planning strategy network proposed in the present invention should focus on traffic participants that are strongly related to the path planning task objectives of the ego vehicle, and predict their driving trajectories in the future to prevent conflicts with the planned path of the ego vehicle. For example, the ego vehicle plans to turn left at a multi-lane intersection without traffic lights. The ego vehicle is located in the first position in the left turn area and there are no other vehicles in front. Then the planning network only needs to pay attention to vehicles that want to enter the target lane, without paying attention to the behavior of other vehicles on the road where the ego vehicle is located. The present invention proposes a network based on a multi-head attention mechanism to capture the ego vehicle's dependence on specific traffic participants, and to satisfy the requirement that the strategy model can accept input vectors that are variable in dimension and arrangement order, because the network structure of the multi-head attention mechanism can ensure that any arrangement of the representation vector input model will obtain the same result, that is, the arrangement of the input vector has order invariance.

[0040] Path planning includes detection, tracking, mapping, motion forecasting, occupancy prediction, and finally safe path planning. The present invention mainly involves tracking and trajectory prediction, which completes the path planning task of the self-vehicle by identifying specific traffic participants related to the self-vehicle path planning task and predicting their and the self-vehicle's motion trajectory.

[0041] See also Figure 1 The present invention provides a vehicle path planning method, comprising:

[0042] Step 100: Obtain image information and related data information of traffic participants and road conditions around the test vehicle through cameras, laser radars, etc.

[0043] The vehicle-mounted camera obtains image information of traffic participants and road conditions around the vehicle. The image information is usually captured from the video information collected by the surround camera, and recognition is performed from the image. For example, the BEV (Bird's-Eye-View) image is input into the target detection model to complete the recognition of traffic participants. Usually, the image is converted into image features, and traffic participants are identified based on the image features. Traffic participants include pedestrians, vehicles, immovable obstacles, etc. The data information of traffic participants around the vehicle includes the position coordinates, wheel camber angle (steering angle), posture state (pitch, yaw, roll), etc. of traffic participants including the vehicle, which can be obtained by radar, etc. The road conditions include the number of lanes on the road where the vehicle is located, the speed limit of each lane, etc., which can be obtained through high-precision maps.

[0044] In one embodiment, the image information and data information of the traffic participants and road conditions around the vehicle are respectively extracted and aggregated using CNN (convolutional neural network) and MLP (multilayer perceptron) to obtain the required feature vector. It should be understood that the information aggregation here is the calculation of all traffic participants around the vehicle, not limited to the target traffic participant.

[0045] Step 200: Determine target traffic participants related to the vehicle path planning task using the image information and data information.

[0046] Through the above steps, all traffic participants around the ego vehicle can be identified. With ego vehicle path planning as the task goal, there will be some traffic participants that do not need to be paid attention to. For example, when the ego vehicle is driving straight in the middle lane, the vehicle behind the ego vehicle in the adjacent lane does not need to be paid attention to; when the ego vehicle changes lanes to the right, the vehicle in the lane to the left of the ego vehicle does not need to be paid attention to; when the ego vehicle changes lanes to the left, the vehicle in the lane to the right of the ego vehicle does not need to be paid attention to.

[0047] Therefore, all traffic participants around the vehicle can be screened according to the current and future planned movement direction of the vehicle, the lane it is in, and the surrounding road conditions, to identify those traffic participants whose future driving trajectories may conflict with the future path planned by the vehicle. In one embodiment, by predicting the future driving trajectory of the traffic participants, the traffic participants whose future driving trajectories are in the target driving lane of the vehicle or the driving lane where the vehicle is located are taken as target traffic participants. That is, the traffic participants whose future driving trajectories may affect the movement direction and lane of the vehicle are taken as calculation objects. It should be understood that the target traffic participants include fixed obstacles in the driving direction or driving lane of the vehicle, and the trajectory of the fixed obstacles is fixed and unchanged.

[0048] After the target traffic participant is identified and determined, the target traffic participant can be tracked based on the intelligent tracking network. Specifically, a set of tracking query vectors can be used to detect the target traffic participant from the image features and continuously track the target traffic participant. By introducing a set of tracking query vectors, the whole process of tracking the target participant from appearing to completely disappearing in the scene is modeled. The tracking query vector is operated by a multi-head cross-attention mechanism with the image features and decoded by a multilayer perceptron (MLP), and finally the boundary position coordinates and velocity, acceleration, angular velocity and other attributes of the tracked object are obtained.

[0049] Since the position coordinates of the traffic participants are known in step 100, the interaction relationship between the target traffic participants and different map elements determined based on the high-precision map generation is used to predict the future trajectory of each target traffic participant. The map generation can be generated by an online map in a 3D scene, and the different categories of map elements in the scene can be segmented, such as lane lines, sidewalks, zebra crossings, dividing lines, drivable areas, etc. (these map elements will be used for the downstream task module to learn the surrounding environment information). Based on the data information and map information input of the target traffic participants around the vehicle, the Actor-Critic composite network based on the attention mechanism can predict the future movement trajectory of the target traffic participants in the scene. Among them, the occupancy grid prediction can use the bird's-eye view feature map as the query vector to determine whether the value of each occupancy grid represents whether the occupancy grid is occupied by the target traffic participant, and the future bird's-eye view feature map is updated through the multi-head attention mechanism, thereby predicting the occupancy grid map for multiple steps in the future.

[0050] Step 300: Utilize a feature extraction model to extract effective feature vectors of the target traffic participant, map the feature vectors into a path point coordinate sequence of the vehicle in the future based on a multi-head attention mechanism and reinforcement learning, and plan a path for the vehicle based on the path point coordinate sequence.

[0051] That is, the above-mentioned CNN (convolutional neural network) and MLP (Multilayer Perceptron) are used for feature extraction to obtain the effective feature vectors of the ego vehicle and the target traffic participant. The actor-critic composite neural network model based on the multi-head-attention mechanism maps the above feature vectors into a sequence of path point coordinates of the ego vehicle in the future as the planned path of the ego vehicle.

[0052] In one embodiment of the present invention, the image information and data information of the traffic participants and road conditions around the vehicle at the current moment and several previous moments are obtained; the image information and data information of the traffic participants and road conditions around the vehicle at the current moment and several previous moments are used as input data of the feature extraction model. By way of example, the present invention uses the image information and data information of the current moment and the 7 moments before the current moment, a total of 8 moments, as the input of the feature extraction network, for example, the perception information is collected every 0.1 seconds, that is, the perception information collected at the current moment and 0.7 seconds before is used. Accordingly, the visual planning path finally output by the present invention is a sequence of the coordinates of the vehicle's path points at the next 8 moments, that is, the function of mapping the next 8 moments into corresponding positions.

[0053] The present invention mainly utilizes CNN and MLP for feature extraction, aggregates the results into a feature vector by splicing, and then inputs it into an Actor-Critic composite network based on the Multi-head-attention mechanism.

[0054] As can be seen from the above, the present invention is a learning-based path planning model, which can determine the target traffic participants related to the path planning task in a complex traffic environment, ignore other irrelevant traffic participants and their behaviors, improve the decision-making ability of the self-vehicle, and make the path planning of the self-vehicle more generalized and flexible; solve the problem that the path planning task flexibility is weak and the path planning is easily affected by changes in the surrounding environment in the prior art. The composite network is composed of an Actor-Critic algorithm framework based on a multi-head attention mechanism. The vehicle is regarded as a single intelligent agent (self-vehicle), and the collected perception information is regarded as the observation space of the intelligent agent. In the Actor-Critic architecture, the output of the Actor part is the action space of the intelligent agent, that is, the path point coordinate sequence of the vehicle in the next 8 moments. The Critic part is the action value function, which is responsible for evaluating the performance of the Actor under a given state s (State) and guiding the Actor's next stage of action. Finally, in the simulation experiment stage, it is only necessary to use the Actor part of the trained model and pass the path point coordinate sequence of the self-vehicle in the next 8 moments output by it to the vehicle motion controller.

[0055] The training process of the above CNN, MLP and Actor-Critic composite network based on Multi-head-attention mechanism is as follows:

[0056] Training data preprocessing:

[0057] In state s0 (state), sensors are used to collect information about traffic participants around the vehicle, and radar and other means are used to obtain data about traffic participants. With the vehicle itself as the center, image information is collected within a range of 50 meters in front, behind, left, and right. It can be seen that the collected area is 100×100m 2 The image is processed by Python's cv2 package to obtain an RGB array, which is then passed to CNN for feature extraction. The tensor finally output by CNN is denoted as c i , c i The dimension of needs to be aligned with the dimension of the feature vector extracted by MLP from the data information of each vehicle. For the data information of N vehicles (including the test vehicle) within a specific range collected by the vehicle sensor, the vector containing the data information of the i-th vehicle is denoted as m i , all m iThe constructed matrices are input into three MLPs for feature extraction, and the query matrix, key matrix and attribute value matrix are obtained. The query vector, key vector and value vector of the i-th vehicle are q i , k i 、v i The aforementioned c i Need to be connected with the q here i 、v i have the same number of elements, c can be used in subsequent calculations i It is regarded as the N+1th vehicle other than the own vehicle.

[0058] Agent training:

[0059] Use the query vector q0 corresponding to the test vehicle and N+1 (one test vehicle plus N surrounding vehicles) key vectors k in sequence i Multiply them together to get the different attention weight vectors a of the tested vehicle for each vehicle i . Note the key vector k i The number of elements of is d, and each weight vector a i Divide by Then use the softmax function to normalize (Normalization) and multiply it by the corresponding v i Finally, the output of the multi-head attention mechanism is summed up and recorded as f (see formula 1), and then the output of the Actor network is obtained through a multi-layer perceptron. The Actor outputs a sequence of path point coordinates for the next five moments.

[0060] ; (Formula 1)

[0061] The path point coordinate sequence output by the Actor part is passed to the vehicle's motion controller. The controller used in the simulation experiment is the PID controller (Proportion Integration Differentiation), and the dynamic model is a two-wheeled bicycle model.

[0062] The agent takes action a0 (action) in state s0. After receiving the agent's action, the environment changes from state s0 to s1, and the environment gives a reward r1 to the agent's behavior. The reward consists of the following parts: whether there is a collision with other traffic participants, the offset between the path point sequence given by the path planner, the offset between the center line of the lane, whether it hits the shoulder, the simulation time, the speed of the vehicle, etc.

[0063] The role of the Critic part is to act as an action-value function (Action-Value Function Q(s,a)), which is used to evaluate the quality of the behavior selected by the Actor. The Critic network consists of three fully connected layers, and the activation function uses the tanh function.

[0064] The network update of Actor-Critic uses the TD3 algorithm (Twin Delayed Deep Deterministic policy gradient algorithm) to update the Actor and Critic network; however, it is not limited to the TD3 algorithm, and other algorithms that conform to the Actor-Critic architecture can also be used. The modification of the network structure of the intelligent agent in the present invention can make it focus on the target traffic participants, ignore the traffic participants that are not related to the path planning task, and effectively combine the information obtained by the intelligent agent from the observation space with the attention mechanism to form the Actor part.

[0065] See also Figure 2 In the present invention, determining the target traffic participant among the traffic participants that is related to the self-vehicle path planning task objective is specifically accomplished by the following steps:

[0066] Step 210: Use a convolutional neural network model and a multi-layer perceptron to extract effective information from the image information and the data information respectively, wherein the convolutional neural network model outputs the extracted feature information as a first feature vector, and the multi-layer perceptron outputs the extracted feature information as a second feature vector. The first feature vector represents the image information, and the second feature vector represents the data information acquired by the sensor.

[0067] The tensor output by CNN for processing image information is denoted as c i , MLP processes the data information and outputs a tensor m containing the data information of each vehicle i That is, CNN outputs the first eigenvector and MLP outputs the second eigenvector.

[0068] Step 220: Align the tensor dimension of the first feature vector with the tensor dimension of the second feature vector; regard the first feature vector as the n+2th second feature vector, input the second feature vector into the multi-head attention mechanism model, and concatenate and aggregate the first feature vector and the second feature vector into a representation tensor of the target traffic participant and the surrounding road conditions for subsequent input into the multi-head attention mechanism network.

[0069] Step 230: Using the representation tensor of the target traffic participant and road condition extracted by the feature extraction model as input, use the Actor part model based on the multi-head attention mechanism and the Actor-Critic algorithm framework in reinforcement learning to predict the path point coordinate sequence of the vehicle in the future, and plan a path for the vehicle based on the path point coordinate sequence.

[0070] After feature extraction through CNN and MLP, the tensor dimensions of the first feature vector and the second feature vector are aligned to realize splicing and aggregation into a feature matrix. That is, the first feature vector is regarded as the second feature vector and input into the subsequent Actor-Critic composite network. The Actor-Critic composite network realizes the identification of the target traffic participant and the path planning of the vehicle by accepting the spliced ​​and aggregated feature matrix as input.

[0071] Among them, the multi-head attention mechanism network is combined with the Actor-Critic algorithm framework, allowing the input tensor to be variable in dimension and arrangement order, and mapping the feature tensor into a sequence of path point coordinates of the vehicle in the future.

[0072] See also Figure 3 The present invention also provides a vehicle path planning device, comprising:

[0073] The acquisition recognition module 31 is used to obtain image information and related data information of traffic participants and road conditions around the test vehicle through cameras, laser radars, etc. CNN (convolutional neural network) and MLP (Multilayer Perceptron) are used for feature extraction and information aggregation to obtain effective feature vectors; the feature vectors are input into the actor-critic composite neural network based on the multi-head-attention mechanism. The composite neural network obtains the image information and data information of traffic participants around the vehicle at the current moment and several previous moments; and the image information and data information of traffic participants around the vehicle at the current moment and several previous moments are used as input data of the feature extraction model. The multi-head attention model combines the Actor-Critic algorithm framework, allows the input tensor to be variable in dimension and arrangement order, and maps the feature vector to a sequence of path point coordinates of the vehicle in the future.

[0074] A target determination module 32, configured to determine a target traffic participant related to a vehicle path planning task using the image information and the data information;

[0075] For example, by predicting the future driving trajectory of the traffic participants, the traffic participants whose future driving trajectories are in the future target driving lane of the vehicle or the driving lane where the vehicle is currently located and conflicting with each other are taken as target traffic participants.

[0076] Specifically, a convolutional neural network model and a multi-layer perceptron are used to extract information from the image information and the data information respectively, the convolutional neural network model outputs the extracted feature information as a first feature vector, and the multi-layer perceptron outputs the extracted feature information as a second feature vector;

[0077] Align the tensor dimension of the first feature vector with the tensor dimension of the second feature vector; wherein the first feature vector represents image information, and the second feature vector represents data information acquired by the sensor; regard the first feature vector as the n+2th second feature vector, and concatenate and aggregate the first feature vector and the second feature vector into a tensor representing the target traffic participant and the surrounding road conditions for subsequent input into the multi-head attention mechanism network.

[0078] The path planning module 33 is used to extract the effective feature vector of the target traffic participant using the feature extraction model, map the feature vector to a path point coordinate sequence of the vehicle in the future based on the multi-head attention mechanism and reinforcement learning, and plan a path for the vehicle based on the path point coordinate sequence. For example, the representation tensor of the target traffic participant and road conditions extracted by the aforementioned feature extraction model is used as input, and the Actor part model based on the multi-head attention mechanism and the Actor-Critic algorithm framework in reinforcement learning is used to predict the path point coordinate sequence of the vehicle in the future, and the path is planned for the vehicle based on the path point coordinate sequence.

[0079] like Figure 4 As shown, the present invention also provides a vehicle control method, comprising:

[0080] Step 410: Obtain image information and related data information of traffic participants and road conditions around the test vehicle through a camera, a laser radar, etc., and identify each traffic participant based on the image information and the data information;

[0081] Step 420: Determine a target traffic participant related to the vehicle path planning task using the image information and the data information;

[0082] Step 430: extracting effective feature vectors of the target traffic participant using a feature extraction model, mapping the feature vectors into a path point coordinate sequence of the vehicle in the future based on a multi-head attention mechanism and reinforcement learning, and planning a path for the vehicle based on the path point coordinate sequence;

[0083] Step 440: Control the vehicle to travel according to the planned vehicle path, for example, by using a PID controller to control the vehicle to travel according to the planned vehicle path.

[0084] The specific content of this embodiment can be referred to the above description and will not be repeated here.

[0085] like Figure 5 As shown, the present invention also provides an electronic device, including:

[0086] At least one processor; and at least one memory in communication with the processor, wherein: the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the above-mentioned vehicle path planning method.

[0087] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned vehicle path planning method is implemented.

[0088] It can be understood that computer-readable storage media may include: any entity or device capable of carrying a computer program, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), and a software distribution medium, etc. A computer program includes a computer program code. The computer program code may be in source code form, object code form, an executable file, or some intermediate form, etc. A computer-readable storage medium may include: any entity or device capable of carrying a computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), and a software distribution medium, etc.

[0089] In certain embodiments of the present invention, the device may include a controller and a processor, wherein the controller is a single-chip microcomputer chip that integrates a processor, a memory, a communication module, etc. The processor may refer to a processor included in the controller. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.

[0090] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

[0091] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A vehicle path planning method, characterized in that: include: Obtain image information and related data information about traffic participants and road conditions around the test vehicle through cameras and lidar; Determining a target traffic participant associated with a vehicle path planning task using the image information and the data information; A feature extraction model is used to extract effective feature vectors of the target traffic participant, the feature vectors are mapped into a path point coordinate sequence of the vehicle in the future based on a multi-head attention mechanism and reinforcement learning, and a path is planned for the vehicle based on the path point coordinate sequence; The extracting the effective feature vector of the target traffic participant by using the feature extraction model includes: Using a convolutional neural network model and a multi-layer perceptron to extract effective information from the image information and the data information respectively, the convolutional neural network model outputs the extracted feature information as a first feature vector, and the multi-layer perceptron outputs the extracted feature information as a second feature vector; the first feature vector and the second feature vector are concatenated into an effective feature vector of the target traffic participant; The step of concatenating the first feature vector and the second feature vector into a valid feature vector of the target traffic participant includes: The tensor dimension of the first feature vector is aligned with the tensor dimension of the second feature vector; the first feature vector is regarded as the n+2th second feature vector, and the first feature vector and the second feature vector are concatenated and aggregated into a representation tensor of the target traffic participant and the surrounding road conditions for subsequent input into a multi-head attention mechanism network.

2. The vehicle path planning method according to claim 1, characterized in that: The method of using the image information and the data information to determine a target traffic participant related to the vehicle path planning task includes: The future driving trajectory of the traffic participant is predicted, and the traffic participant whose future driving trajectory is in the target driving lane of the vehicle or in the driving lane where the vehicle is located is taken as the target traffic participant.

3. The vehicle path planning method according to claim 1, characterized in that: The first feature vector represents valid image information extracted from the camera, and the second feature vector represents valid data information extracted from the sensor.

4. The vehicle path planning method according to claim 3, characterized in that: The multi-head attention mechanism is combined with the Actor-Critic algorithm architecture in reinforcement learning, allowing the input tensor to be variable in dimension and arrangement order; the number of target traffic participants around the test vehicle is variable and the arrangement order of the representation vectors of each target traffic participant is variable; and the feature vector can be mapped into a sequence of path point coordinates of the vehicle in the future.

5. The vehicle path planning method according to claim 2, characterized in that: The method of using the image information and the data information to determine a target traffic participant related to the vehicle path planning task includes: Obtain image information and data information of traffic participants around the test vehicle at the current moment and several previous moments; use the image information and data information of traffic participants around the vehicle at the current moment and several previous moments as input data for the feature extraction model.

6. A vehicle path planning device, characterized in that: include: Acquisition and recognition module, used to obtain image information and relevant data information of traffic participants and road conditions around the test vehicle through cameras and lidar; A target determination module, used to determine a target traffic participant related to a vehicle path planning task using the image information and the data information; A path planning module is used to extract effective feature vectors of the target traffic participant using a feature extraction model, map the feature vectors into a path point coordinate sequence of the vehicle in the future based on a multi-head attention mechanism and reinforcement learning, and plan a path for the vehicle based on the path point coordinate sequence; The path planning module is also used to use a convolutional neural network model and a multi-layer perceptron to extract effective information from the image information and the data information respectively, the convolutional neural network model outputs the extracted feature information as a first feature vector, and the multi-layer perceptron outputs the extracted feature information as a second feature vector; the first feature vector and the second feature vector are spliced ​​into a valid feature vector of the target traffic participant; and the tensor dimension of the first feature vector is aligned with the tensor dimension of the second feature vector; the first feature vector is regarded as the n+2th second feature vector, and the first feature vector and the second feature vector are spliced ​​and aggregated into a representation tensor of the target traffic participant and the surrounding road conditions for subsequent input into a multi-head attention mechanism network.

7. A vehicle control method, characterized in that: include: Obtain a vehicle planned path determined by the vehicle path planning method according to any one of claims 1 to 5, and control vehicle driving according to the vehicle planned path.

8. An electronic device, characterized in that: include: at least one processor; And at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method as described in any one of claims 1 to 5 and 7.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a computer, the method according to any one of claims 1 to 5 and 7 is executed.

Images