Track planning method and device, vehicle and computer readable storage medium

Abstract

The invention provides a trajectory planning method and device, a vehicle and a computer readable storage medium, and the method comprises the steps: obtaining vehicle driving information, and generating a multi-modal trajectory according to the vehicle driving information, the multi-modal trajectory comprising a plurality of sub-trajectories; obtaining a hierarchical strategy network model, and inputting the multi-modal trajectory and the vehicle driving information into the hierarchical strategy network model to obtain a multi-dimensional score corresponding to each sub-trajectory; and determining a target trajectory in the plurality of sub-trajectories according to the multi-dimensional score. By setting a hierarchical strategy network model comprising a plurality of strategy layers, sub-tracks can be scored based on different scoring dimensions through different strategy layers, so that target tracks meeting different dimension requirements can be determined by integrating multi-dimensional driving requirements; therefore, the obtained target trajectory can meet vehicle driving requirements and adapt to a dynamic driving environment in an actual scene.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving, and in particular to a trajectory planning method, device, vehicle, and computer-readable storage medium. Background Art

[0002] Autonomous driving technology has made significant progress in recent years. Trajectory planning is a core issue in autonomous driving. Trajectories in existing technologies are often planned based on routes to the destination, resulting in rigid decision-making and difficulty adapting to the dynamic driving environment in actual driving scenarios. Summary of the Invention

[0003] The main purpose of the present invention is to propose a trajectory planning method, device, vehicle and computer-readable storage medium, aiming to solve the problem that trajectory planning in the prior art is difficult to adapt to the actual driving environment.

[0004] To achieve the above object, the present invention provides a trajectory planning method, which includes the following steps:

[0005] Acquiring vehicle driving information, and generating a multimodal trajectory based on the vehicle driving information, wherein the multimodal trajectory includes a plurality of sub-trajectories;

[0006] Obtaining a hierarchical policy network model, and inputting the multimodal trajectory and the vehicle driving information into the hierarchical policy network model to obtain a multidimensional score corresponding to each sub-trajectory, wherein the hierarchical policy network model includes multiple policy layers, each policy layer scores each sub-trajectory based on different scoring rules combined with the vehicle driving information, and the multidimensional score is comprehensively determined by the scores of each policy layer;

[0007] A target track is determined from the plurality of sub-tracks according to the multi-dimensional scores.

[0008] Optionally, the hierarchical strategy network model includes a macro-strategy layer, a meso-strategy layer, and a micro-strategy layer; inputting the multimodal trajectory and the vehicle driving information into the hierarchical strategy network model to obtain a multi-dimensional score corresponding to each sub-trajectory includes:

[0009] Obtaining target navigation information from the vehicle driving information, and scoring the sub-trajectory according to the target navigation information by the macro-strategy layer to obtain a macro score, wherein the macro score indicates the degree to which the sub-trajectory conforms to the navigation information;

[0010] Obtaining environmental information from the vehicle driving information, and scoring the sub-trajectory according to the environmental information through the meso-strategy layer to obtain a meso-score, wherein the meso-score indicates the legality of the interaction between the sub-trajectory and the environmental information;

[0011] Obtaining a preset trajectory standard, and scoring the sub-trajectory according to the preset trajectory standard by the micro-strategy layer to obtain a micro-score, wherein the micro-score indicates the degree to which the sub-trajectory meets the preset trajectory standard;

[0012] The multi-dimensional score is obtained by integrating the macro score, the meso score and the micro score.

[0013] Optionally, the scoring of the sub-trajectory by the meso-strategy layer according to the environmental information to obtain a meso-score, wherein the meso-score indicating the legality of the interaction between the sub-trajectory and the environmental information includes:

[0014] Obtaining the predicted trajectory of the surrounding vehicle and characteristic object information in the environmental information;

[0015] Determining interactive predictions between the sub-trajectory, the surrounding vehicle predicted trajectory, and the feature object;

[0016] The macro score is obtained, and the meso score is determined based on the interactive prediction and the macro score.

[0017] Optionally, the scoring of the sub-trajectory according to the preset trajectory standard by the micro-strategy layer to obtain a micro-score, wherein the micro-score indicates the degree to which the sub-trajectory meets the preset trajectory standard, includes:

[0018] Obtaining a kinematic standard and a comfort standard in the preset trajectory standard;

[0019] Each of the sub-trajectories is scored according to the kinematic standard and the comfort standard to obtain the micro-score.

[0020] Optionally, determining a target track from the plurality of sub-tracks according to the multi-dimensional scores includes:

[0021] Obtaining an intermediate score corresponding to the sub-track according to the macro score and the meso score;

[0022] Determining a plurality of high-ranking sub-trajectories in the sub-trajectories, wherein the high-ranking sub-trajectories are a preset number of the sub-trajectories with correspondingly higher intermediate scores;

[0023] Determine whether the microscopic scores corresponding to the high-order sub-trajectories all reach a preset microscopic threshold;

[0024] If the micro scores corresponding to all the high-level sub-trajectories are less than the preset micro threshold, an adjustment instruction is sent to the macro-strategy layer to cause the macro-strategy layer to reduce the macro scores corresponding to the high-level sub-trajectories to update the macro scores;

[0025] If there is a microscopic score corresponding to the high-order sub-trajectory that is greater than a preset microscopic threshold, the target trajectory is determined in the high-order sub-trajectory according to the multi-dimensional score.

[0026] Optionally, generating a multimodal trajectory according to the vehicle driving information includes:

[0027] Obtaining an expected driving route based on the vehicle driving information path planning;

[0028] Determining a plurality of target points on the expected driving route, and generating a plurality of anchor point trajectories based on the current position of the vehicle and the target points;

[0029] For each of the anchor point trajectories, performing noise superposition on the anchor point trajectories to obtain multiple noise trajectories corresponding to the anchor point trajectories;

[0030] The anchor point trajectory and the noise trajectory are combined to obtain the multimodal trajectory.

[0031] Optionally, the combining the anchor point trajectory and the noise trajectory to obtain the multimodal trajectory includes:

[0032] determining a trajectory driving intention corresponding to each of the anchor point trajectories and the noise trajectory;

[0033] Obtaining a target driving intention corresponding to the expected driving route;

[0034] Determine the trajectory driving intention and the target driving intention in the anchor point trajectory and the noise trajectory Figure 1 The corresponding sub-trajectory;

[0035] The sub-trajectories are aggregated to obtain the multimodal trajectory.

[0036] To achieve the above object, the present invention further provides a trajectory planning device, the trajectory planning device comprising:

[0037] a first acquisition module, configured to acquire vehicle driving information and generate a multimodal trajectory according to the vehicle driving information, wherein the multimodal trajectory includes a plurality of sub-trajectories;

[0038] a second acquisition module, configured to acquire a hierarchical strategy network model, and input the multimodal trajectory and the vehicle driving information into the hierarchical strategy network model to obtain a multidimensional score corresponding to each sub-trajectory, wherein the hierarchical strategy network model includes multiple strategy layers, each strategy layer scores each sub-trajectory based on different scoring rules combined with the vehicle driving information, and the multidimensional score is determined comprehensively by the scores of each strategy layer;

[0039] A first determination module is configured to determine a target track from the plurality of sub-tracks according to the multi-dimensional scores.

[0040] To achieve the above objectives, the present invention also provides a vehicle, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the trajectory planning method described above when executed by the processor.

[0041] To achieve the above objectives, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the trajectory planning method described above are implemented.

[0042] The present invention proposes a trajectory planning method, device, vehicle, and computer-readable storage medium. The method obtains vehicle driving information and generates a multimodal trajectory based on the vehicle driving information, wherein the multimodal trajectory includes multiple sub-trajectories. The method also obtains a hierarchical strategy network model and inputs the multimodal trajectory and the vehicle driving information into the hierarchical strategy network model to obtain a multidimensional score corresponding to each sub-trajectory. The hierarchical strategy network model includes multiple strategy layers, each of which scores each sub-trajectory based on different scoring rules combined with the vehicle driving information. The multidimensional score is determined by comprehensively combining the scores of each strategy layer. A target trajectory is determined from the multiple sub-trajectories based on the multidimensional scores. By setting up a hierarchical strategy network model comprising multiple strategy layers, the sub-trajectories can be scored based on different scoring dimensions using different strategy layers. This allows the target trajectory to be determined based on the multidimensional driving requirements, thereby comprehensively determining a target trajectory that meets the vehicle driving requirements and adapts to the dynamic driving environment in actual scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0045] Figure 1 This is a flow chart of the first embodiment of the trajectory planning method of the present invention;

[0046] Figure 2 A detailed flow chart of the trajectory planning method of the present invention;

[0047] Figure 3 Schematic diagram of the process of the hierarchical strategy network model in the trajectory planning method of the present invention;

[0048] Figure 4 Schematic diagram of the module structure of the vehicle of the present invention. DETAILED DESCRIPTION

[0049] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only embodiments of a part of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of this application.

[0050] The present invention provides a trajectory planning method, referring to Figure 1 , Figure 1 This is a flow chart of a first embodiment of a trajectory planning method according to the present invention, wherein the method comprises the following steps:

[0051] Step S10: Acquire vehicle driving information, and generate a multimodal trajectory based on the vehicle driving information, wherein the multimodal trajectory includes a plurality of sub-trajectories;

[0052] Vehicle driving information is relevant information during the vehicle's driving process; vehicle driving information can specifically include information about the vehicle itself and information about the driving scene. For example, information about the vehicle itself can include vehicle position, speed, head orientation, navigation information, etc., and information about the driving scene can include images collected by cameras, point cloud data collected by lidar, etc.; the specific data collection method can be set based on actual needs, such as setting a multi-perspective camera on the vehicle body, and collecting multi-perspective images of the environment around the vehicle through the multi-perspective camera. At the same time, the multi-perspective image can be feature extracted through the image coding network to determine the relevant information in the multi-perspective image, such as surrounding vehicles, railings, pedestrians, traffic lights, road signs, road structures, etc. contained in the multi-perspective image; the multi-perspective image can also be spatially transformed to uniformly map images of different perspectives to a bird's-eye view. A bird's-eye view is obtained to achieve a global perception of the vehicle's surrounding environment; a lidar is installed on the vehicle body to obtain point cloud data. By extracting features from the point cloud data, relevant information in the environment, such as obstacles, curbs, lane lines, etc., can be identified. At the same time, the relevant information of the point cloud recognition can be converted to the same spatial dimension of the multi-view image, thereby achieving alignment of visual and spatial information, thereby making full use of the advantages of point clouds in spatial structure and distance perception, and making up for the shortcomings of image information in depth estimation and occlusion processing; point cloud data can also be integrated with multi-view images to obtain driving scene information. The driving scene information can improve the accuracy and robustness of environmental modeling, and can provide rich and accurate environmental input for the subsequent trajectory planning model, so as to better understand and respond to complex and changing traffic scenes.

[0053] A multimodal trajectory consists of multiple sub-trajectories with different states. Each sub-trajectory generates multiple preliminary feasible trajectories corresponding to the vehicle's current driving state. The specific method for determining sub-trajectories can be set based on actual needs. For example, the vehicle's expected driving route can be first obtained through navigation information, and then multiple sub-trajectories that can be driven based on the expected driving route are determined based on the expected driving route.

[0054] Step S20: Obtain a hierarchical strategy network model, and input the multimodal trajectory and the vehicle driving information into the hierarchical strategy network model to obtain a multidimensional score corresponding to each sub-trajectory, wherein the hierarchical strategy network model includes multiple strategy layers, each strategy layer scores each sub-trajectory based on different scoring rules combined with the vehicle driving information, and the multidimensional score is determined by comprehensively determining the scores of each strategy layer;

[0055] Step S30 : determining a target track from the plurality of sub-tracks according to the multi-dimensional scores.

[0056] The hierarchical strategy network model is configured with multiple strategy layers, each of which scores sub-trajectories. Specifically, each strategy layer uses different scoring rules for scoring sub-trajectories, and the scoring rules corresponding to different strategy layers can be set based on actual needs. It can be understood that the scoring rules are set based on vehicle safety, compliance, comfort, and efficiency. Therefore, the scoring of sub-trajectories using the scoring rules should be able to reflect the positive effects of the sub-trajectory on vehicle driving. For example, when a higher score indicates that the sub-trajectory is more likely to be determined as the target trajectory, if the sub-trajectory can make vehicle driving safer, the corresponding score of the sub-trajectory can be higher; if the sub-trajectory can make vehicle driving more compliant, the corresponding score of the sub-trajectory can be higher; if the sub-trajectory can make vehicle driving more comfortable, the corresponding score of the sub-trajectory can be higher; if the sub-trajectory can make vehicle driving more efficient, the corresponding score of the sub-trajectory can be higher. Different strategy layers can be configured to target the different impacts of sub-trajectories on vehicle driving. For example, the first strategy layer sets scoring rules based on the safety of the sub-trajectory, the second strategy layer sets scoring rules based on whether the sub-trajectory complies with traffic regulations, the third strategy layer sets scoring rules based on the smooth driving of the vehicle caused by the sub-trajectory, and the fourth strategy layer sets scoring rules based on the speed of the vehicle under the sub-trajectory. It should be noted that in actual applications, sub-trajectories can be scored positively (i.e., the higher the score, the more likely the sub-trajectory will be selected as a target trajectory), or negatively (i.e., the lower the score, the more likely the sub-trajectory will be selected as a target trajectory). This embodiment and subsequent embodiments use positive scoring as an example, and negative scoring can be implemented by analogy and will not be further described.

[0057] It is understandable that the vehicle driving information reflects the information of the vehicle itself and the environment. Therefore, by combining the vehicle driving information, the influence of the sub-trajectory on the vehicle in actual driving can be further determined. For example, when scoring the safety of the sub-trajectory, the relevant information of the surrounding vehicles in the vehicle driving information can be used to determine whether the self-vehicle will collide with the surrounding vehicles when driving on the sub-trajectory, thereby achieving safety scoring. When setting specific scoring rules, the distance between the self-vehicle and the surrounding vehicles can be used as a scoring reference. For example, the smaller the distance between the self-vehicle and the surrounding vehicles, the lower the corresponding score; for example, when scoring the compliance of the sub-trajectory, the distance between the self-vehicle and the surrounding vehicles can be used as a scoring reference. When scoring, the traffic information, road signs, road structure, etc. in the vehicle driving information can be used to determine whether the sub-trajectory will cause traffic violations, thereby achieving compliance scoring. When setting specific scoring rules, the presence of illegal behavior can be used as a scoring reference. For example, if the road signs include solid white lines, the distance between the vehicle and the solid white line can be set as a scoring reference. The smaller the distance between the vehicle and the solid white line, the lower the corresponding score. When the vehicle changes lanes across the solid white line, the corresponding score is 0. The application of other types of vehicle driving information in scoring can also be implemented in an analogous way and will not be elaborated here.

[0058] The specific number of strategy layers and the scoring rules for each strategy layer can be set based on actual application scenarios and needs.

[0059] The multi-dimensional score is obtained by the scores output by each policy layer in the hierarchical policy network model. The specific calculation method of the multi-dimensional score can be set based on actual needs, such as taking the sum of the scores corresponding to each policy layer as the multi-dimensional score; the scores corresponding to each policy layer can also be weighted averaged, and different policy layers are weighted based on the importance or influence of their scoring rules on vehicle driving. The higher the importance or influence, the greater the corresponding weight.

[0060] By combining driving information with a hierarchical policy network model containing multiple policy layers, a multidimensional score is generated. The multidimensional score can comprehensively evaluate sub-trajectories based on the actual driving environment and the different focuses of vehicle driving, and comprehensively determine the impact of sub-trajectories on vehicle driving. Based on the multidimensional score, the target trajectory that best suits the current vehicle driving scenario can be selected from the multimodal sub-trajectories.

[0061] This embodiment establishes a hierarchical policy network model consisting of multiple policy layers, enabling different policy layers to score sub-trajectories based on different scoring dimensions. This allows the target trajectory to be determined based on the multi-dimensional driving requirements, ensuring that the resulting target trajectory meets the vehicle's driving requirements and adapts to the dynamic driving environment in real-world scenarios.

[0062] Further, see Figure 2 In the second embodiment of the trajectory planning method of the present invention proposed based on the first embodiment of the present invention, the hierarchical strategy network model includes a macro-strategy layer, a meso-strategy layer, and a micro-strategy layer; step S20 includes the following steps:

[0063] Step S21, obtaining target navigation information from the vehicle driving information, and scoring the sub-trajectory according to the target navigation information by the macro-strategy layer to obtain a macro score, wherein the macro score indicates the degree to which the sub-trajectory conforms to the navigation information;

[0064] The macro-strategy layer is used to evaluate global path planning and long-distance decision-making; therefore, the macro-strategy layer scores the sub-trajectories based on the target navigation information in the vehicle driving information; the macro score is the scoring result output by the macro-strategy layer.

[0065] The target navigation information is the relevant information for navigation output by the vehicle's navigation system; it can be understood that the target navigation information indicates the future relevant planning of the vehicle during driving, so the evaluation of global path planning and long-distance decision-making can be achieved through the target navigation information; the specific parameters contained in the target navigation information can be set based on actual needs, such as the target navigation information can specifically include the expected driving route, optional lanes, target speed, traffic flow adaptation factor; the expected driving route is the driving route planned by navigation, if the driving end point has been determined, the navigation system will plan the route to the driving end point based on the current position of the vehicle to obtain the expected driving route; the optional lanes are the vehicle on the expected driving route. Lanes that can be driven. For example, if a vehicle is about to pass through an intersection and the expected driving route indicates that it needs to go straight, the corresponding optional lanes are straight lanes, left-turn straight lanes, and right-turn straight lanes. At the same time, left-turn lanes and right-turn lanes are prohibited lanes; the target speed is the required speed during driving. The target speed can be set to the maximum speed, minimum speed, recommended speed or speed range. For example, if the maximum speed required for the lane where the current vehicle is located is 60, the target speed can be set to the maximum speed of 60; the traffic flow adaptation factor is used to indicate traffic volume, traffic speed, vehicle type, etc. The overall state of traffic can be known through the traffic flow factor. By setting the traffic flow factor, the path can be dynamically adjusted to avoid congestion and high-risk areas.

[0066] The specific scoring rules of the macro-strategy layer can be set based on actual needs. Taking the above-mentioned target navigation information as an example, scoring rules can be set for optional lanes and target speeds respectively. For example, for optional lanes, it is determined that the vehicle can stay in the optional lane under the sub-trajectory. When the vehicle always stays in the optional lane, the sub-trajectory is considered to be more consistent with the navigation information and is scored higher. When the vehicle exceeds the optional lane and reaches the no-entry lane, the sub-trajectory is considered to be inconsistent with the navigation information and is scored lower, 0, or a penalty point is given. The penalty point is negative. By assigning penalty points to sub-trajectories with major problems, the multi-dimensional scores corresponding to the sub-trajectory can be greatly reduced, thereby avoiding sub-trajectories with major problems as target trajectories as much as possible. For example, for target speed, when the vehicle speed corresponding to the sub-trajectory exceeds the maximum speed or is lower than the minimum speed, a penalty point can be assigned.

[0067] Macro scoring can be directly obtained by integrating the evaluation results of multiple parameters in the target navigation information, such as:

[0068] a macro =(lane_choice,target_speed,α traffic )

[0069] Among them, a macrois the macro score; lane_choice is the score corresponding to the optional lane; target_speed is the score corresponding to the target speed; α traffic is the score corresponding to the traffic flow adaptation factor.

[0070] It can also be obtained by scoring based on specific parameters such as optional lanes and target speed, and then comprehensively obtaining the result, such as using the total score as a macro score or performing a weighted average.

[0071] Macro scoring scores sub-trajectories based on navigation information, so that the target trajectory obeys the planned navigation situation and ensures that the target trajectory can meet the basic correct route driving.

[0072] Step S22: obtaining environmental information from the vehicle driving information, and scoring the sub-trajectory according to the environmental information through the meso-strategy layer to obtain a meso-score, wherein the meso-score indicates the legality of the interaction between the sub-trajectory and the environmental information;

[0073] The meso-strategy layer is used to evaluate the interaction between the vehicle and the environment; therefore, the meso-strategy layer scores the sub-trajectories based on the environmental information in the vehicle driving information. The meso-score is the scoring result output by the meso-strategy layer.

[0074] Environmental information is relevant information about the environment in which the vehicle is located. It can be understood that environmental information indicates the relative situation of the vehicle and the environment during driving. Therefore, environmental information can be used to evaluate the interaction between the vehicle and the environment; the specific parameters contained in the environmental information can be set based on actual needs. For example, environmental information can include the predicted trajectory of the surrounding vehicle, lane lines, curbs, guardrails, sidewalks, etc.

[0075] The specific scoring rules of the meso-strategy layer can be set based on actual needs. Taking the above-mentioned environmental information as an example, for the trajectory, curb or guardrail predicted by the surrounding vehicle, it is determined whether the sub-trajectory will collide with the trajectory, curb or guardrail predicted by the surrounding vehicle. If a collision occurs, the sub-trajectory is considered illegal and a penalty point is assigned. If no collision occurs, the sub-trajectory is considered legal, and the score of the sub-trajectory corresponding to the farther distance from the surrounding vehicle can be assigned a higher score; for lane lines, sidewalks, etc., they can be set based on traffic rules. For example, if the lane line is a solid white line, the distance between the ego vehicle and the solid white line can be set as a scoring reference. The smaller the distance between the ego vehicle and the solid white line, the lower the corresponding score. When the ego vehicle changes lanes through the solid white line, the sub-trajectory is considered illegal, and the corresponding score is 0 or a penalty point is set.

[0076] In specific implementation, the vehicle and the environment can be dynamically modeled, and the relationship between the vehicle and other objects in the environment can be determined through dynamic modeling. In order to further clarify the relationship between the vehicle and the environment, the vehicle's own information, such as vehicle position, speed, and head direction, can also be obtained and modeled in combination with the vehicle's own information. The specific dynamic modeling method can be set based on actual needs, such as using MARL (Multi-Agent RL) and GNN (Graph Neural Network) to combine dynamic modeling. Through graph neural networks, efficient information flow and decision-making coordination can be achieved, thereby improving safety and game-playing capabilities in multi-vehicle scenarios.

[0077] Meso-level scoring can be directly obtained by integrating the evaluation results of multiple parameters of environmental information, such as:

[0078] a meso =(heading,obstacle_avoidance, )

[0079] Among them, a meso is the mesoscopic score; heading is the score corresponding to the yaw angle; obstacle_avoidance is the obstacle avoidance score between the feature object; is the collision probability with the Nth peripheral vehicle.

[0080] It can also be obtained based on specific parameters such as scoring and then synthesizing them, such as using the total score as a meso-level score or taking a weighted average.

[0081] Meso-level scoring evaluates sub-trajectories through the interaction between the vehicle and the environment, thereby ensuring the safety and compliance of sub-trajectories and avoiding collisions or violations.

[0082] Step S23: obtaining a preset trajectory standard, and scoring the sub-trajectory according to the preset trajectory standard by the micro-strategy layer to obtain a micro-score, wherein the micro-score indicates the degree to which the sub-trajectory meets the preset trajectory standard;

[0083] The micro-strategy layer is used to evaluate the rationality and comfort of the sub-trajectory itself; therefore, the micro-strategy layer scores the sub-trajectory based on the preset trajectory standards; the micro-score is the scoring result output by the micro-strategy layer.

[0084] The preset trajectory standard is relevant information for judging the rationality and comfort of the trajectory; the specific parameters contained in the preset trajectory standard can be set based on actual needs, such as the preset trajectory standard can specifically include kinematic laws and vehicle comfort requirements; kinematic laws are the movement characteristics of the vehicle, such as the vehicle's acceleration performance and steering angular rate performance. It can be understood that the kinematic laws are set based on the actual capabilities of the vehicle. Therefore, the feasibility of the sub-trajectory can be determined through the kinematic laws; the vehicle comfort requirement is the user's comfort level under the relevant parameters of the vehicle. For example, if the vehicle is subject to sudden acceleration, sudden deceleration, or steering wheel shaking, the body of the vehicle will shake, resulting in poor user comfort.

[0085] The specific scoring rules at the micro-strategy layer can be set based on actual needs. Taking the above-mentioned preset trajectory standards as an example, scoring rules can be set for kinematic laws and vehicle comfort requirements respectively. For example, for kinematic laws, the acceleration and steering angular rate required for the vehicle under the sub-trajectory are determined. If the acceleration exceeds the maximum speed of the kinematic law or the steering angular rate exceeds the maximum steering angular rate of the kinematic law, the sub-trajectory is considered to not meet the preset trajectory standards and is scored as 0 or given a penalty point. For example, for vehicle comfort, comfort thresholds can be set for acceleration, deceleration, and steering wheel vibration. Taking acceleration as an example, if the acceleration exceeds the comfort threshold, the sub-trajectory is considered to be uncomfortable and does not meet the preset trajectory standards. Therefore, it is scored as 0 or given a penalty point. If the acceleration is less than the comfort threshold, the smaller the acceleration, the higher the corresponding score.

[0086] The micro-score can be directly obtained by integrating the evaluation results of multiple parameters within the preset trajectory standards, or it can be obtained by scoring specific parameters and then integrating them, such as using the total score as the micro-score or taking a weighted average.

[0087] Step S24, combining the macro score, the meso score and the micro score to obtain the multi-dimensional score.

[0088] After determining the macro, meso, and micro scores, a comprehensive multidimensional score for the sub-trajectory can be obtained. It is understandable that the multidimensional score is determined based on the comprehensive determination of different strategic layers. Therefore, the multidimensional score can integrate the scoring rules of different strategic layers, making it possible to evaluate the sub-trajectory in multiple aspects such as global path planning and long-distance decision-making, interaction between the vehicle and the environment, trajectory rationality and comfort, etc., thereby obtaining a target trajectory that meets the requirements in all dimensions.

[0089] The specific parameters of the hierarchical policy network can be set based on actual needs, such as reward function and model optimization.

[0090] When setting the reward function, in addition to traditional progress, safety, comfort, rules, and efficiency parameters, adaptive weights and risk scheduling mechanisms can also be introduced.

[0091] Using the Meta-Learning method, the reward weights of different parameters are adaptively optimized during the training process:

[0092]

[0093] Among them, λ i is the reward weight of the i-th parameter, π is the joint layered strategy, and R indicates the reward corresponding to each layer strategy.

[0094] Constructing a risk scheduling mechanism by introducing risk penalty items:

[0095]

[0096] Where t indicates the trajectory number; R risk is the risk penalty term; Var[τ t ] is the trajectory prediction uncertainty; δ is the risk threshold; uncertainty(t) is the confidence level.

[0097] Combining the adaptive weight and risk scheduling mechanism, we can get the final reward function:

[0098] R t =λ1R progress +λ2R safety +λ3R comfort +λ4R rules +λ5R efficiency +λ6R risk

[0099] Among them, R progress is the reward item corresponding to the progress; R safety is the reward item corresponding to safety; R comfort R is the reward item corresponding to comfort; rules is the reward item corresponding to the rule; R efficiency It is the reward item corresponding to efficiency.

[0100] When optimizing the model, you can first configure the simulation environment, such as using CARLA 2.0+

[0101] MetaDrive is used as a simulation platform, and the scene complexity is set to a dynamic obstacle density of 0.5–2.0 obstacles per 100 meters; the interactive objects are set to pedestrians, vehicles, and special traffic participants.

[0102] Policy Gradient Optimization:

[0103] Adopting the PPO (Proximal Policy Optimization) framework for hierarchical updates:

[0104] Macro strategy update:

[0105]

[0106] A GAE is the advantage function, which is calculated by accumulating 20 trajectories; a is the score predicted by the macro-level model; S represents the trajectory; a|s represents the score obtained given the trajectory s.

[0107] Furthermore, the step S22 includes the steps of:

[0108] Step S221, obtaining the predicted trajectory of the surrounding vehicle and the characteristic object information in the environmental information;

[0109] Step S222, determining the interactive prediction between the sub-trajectory, the surrounding vehicle predicted trajectory, and the feature object;

[0110] Step S223: Obtain the macro score, and determine the meso score based on the interaction prediction and the macro score.

[0111] Feature information indicates relevant entity information in the environment, such as lane lines, curbs, guardrails, sidewalks, etc.

[0112] Interaction prediction indicates the possible interactions between the vehicle and surrounding vehicles or features in the environment, such as collision, lane change, etc.

[0113] In this embodiment, the macro-scoring is combined with the comprehensive scoring in the meso-strategy layer. It can be understood that the macro-strategy layer and the meso-strategy layer score the sub-trajectory according to different dimensions. Specifically, the macro-strategy layer scores in the dimensions of global path planning and long-distance decision-making, and the meso-strategy layer scores in the dimension of interaction between the vehicle and the environment. Therefore, there may be two strategy layers that give relatively different scores for the same sub-trajectory, and after combining the macro-scoring and the meso-strategy scoring, the score of the sub-trajectory may be higher, while the problems of the sub-trajectory are hidden. For example, in the following vehicle scenario, if trajectory A is following the vehicle in front, trajectory B is changing lanes to the left to overtake, and trajectory C is changing lanes to the right to overtake; judging only from the interaction with the surrounding vehicles, that is, the overall scoring rules of the meso-strategy layer, the three are all reasonable trajectories, and the specific scoring situation will not cause a large gap; however, if the navigation information indicates preparing to exit the ramp at this time, then trajectory B is obviously unreasonable, because if you change lanes to the left to overtake, you will miss the ramp; if the lane you are in at this time is a long straight and the minimum speed limit is 80, but the speed of the car in front is 40, then trajectory A is unreasonable, and following the car cannot make the speed reach the minimum speed limit; therefore, in order to further improve the accuracy of the meso-level scoring, in this embodiment, the meso-level strategy layer is combined with the macro-level scoring for comprehensive scoring. It can be understood that the macro-level scoring will include scores and related data corresponding to different information, such as information including optional lanes and target elements. The optional lanes can indicate preparing to exit the ramp or the long straight, and the target speed can be only the minimum speed limit of 80. Therefore, from the macro-level scoring, the meso-level strategy layer can obtain relevant information at the macro level, thereby realizing comprehensive scoring in combination with the macro-level scoring.

[0114] Furthermore, the step S23 includes the steps of:

[0115] Step S231, obtaining the kinematic standard and the comfort standard in the preset trajectory standard;

[0116] Step S232 : scoring each of the sub-trajectories according to the kinematics standard and the comfort standard to obtain the microscopic score.

[0117] In this embodiment, the preset trajectory standards specifically include kinematic laws and vehicle comfort requirements. Kinematic laws refer to the vehicle's motion characteristics, such as its acceleration performance and steering angular rate performance. It is understandable that the kinematic laws are set based on the vehicle's actual capabilities. Therefore, the kinematic laws can be used to determine whether a sub-trajectory is feasible. The vehicle comfort requirement refers to the user's comfort level under the relevant vehicle parameters. For example, rapid acceleration, rapid deceleration, and steering wheel vibration may cause the vehicle body to shake, resulting in poor user comfort.

[0118] The specific scoring rules at the micro-strategy layer can be set based on actual needs. Taking the above-mentioned preset trajectory standards as an example, scoring rules can be set for kinematic laws and vehicle comfort requirements respectively. For example, for kinematic laws, the acceleration and steering angular rate required for the vehicle under the sub-trajectory are determined. If the acceleration exceeds the maximum speed of the kinematic law or the steering angular rate exceeds the maximum steering angular rate of the kinematic law, the sub-trajectory is considered to not meet the preset trajectory standards and is scored as 0 or given a penalty point. For example, for vehicle comfort, comfort thresholds can be set for acceleration, deceleration, and steering wheel vibration. Taking acceleration as an example, if the acceleration exceeds the comfort threshold, the sub-trajectory is considered to be uncomfortable and does not meet the preset trajectory standards. Therefore, it is scored as 0 or given a penalty point. If the acceleration is less than the comfort threshold, the smaller the acceleration, the higher the corresponding score.

[0119] The micro-score can be directly obtained by integrating the evaluation results of multiple parameters within the preset trajectory standards, or it can be obtained by scoring specific parameters and then integrating them, such as using the total score as the micro-score or taking a weighted average.

[0120] Through micro-scoring, the rationality and comfort of the sub-trajectory itself can be directly evaluated, thereby determining whether the sub-trajectory itself can be reasonably implemented and ensuring the control quality of the sub-trajectory itself.

[0121] Further, see Figure 3 In the third embodiment of the trajectory planning method of the present invention proposed based on the first embodiment of the present invention, step S30 includes the following steps:

[0122] Step S31, obtaining an intermediate score corresponding to the sub-track according to the macro score and the meso score;

[0123] Step S32, determining a plurality of high-ranking sub-trajectories in the sub-trajectories, wherein the high-ranking sub-trajectories are a preset number of the sub-trajectories with corresponding higher intermediate scores;

[0124] Step S33, determining whether the microscopic scores corresponding to the high-order sub-trajectories all reach a preset microscopic threshold;

[0125] In step S34, if the micro scores corresponding to all the high-level sub-trajectories are less than the preset micro threshold, an adjustment instruction is sent to the macro-strategy layer to cause the macro-strategy layer to reduce the macro scores corresponding to the high-level sub-trajectories to update the macro scores.

[0126] Step S35 : If there is a high-ranking sub-trajectory whose corresponding microscopic score is greater than a preset microscopic threshold, the target trajectory is determined in the high-ranking sub-trajectory according to the multi-dimensional score.

[0127] The intermediate score is a comprehensive score obtained by combining the macro score and the meso score.

[0128] Sort each sub-track according to the middle score from high to low, and the preset number of sub-tracks before the high school entrance examination are ranked as high-ranking sub-tracks; the specific value of the preset number can be set based on actual needs; it can be set to a specific value, such as 100, or it can be set to a proportion, such as 10% of the number of sub-tracks.

[0129] It is understandable that the micro-scoring is to evaluate the rationality and comfort of the sub-trajectory itself.

[0130] Therefore, when the micro-score is low, it means that the sub-track itself is not suitable for application.

[0131] The preset micro-threshold is used to indicate that a sub-trajectory is unreasonable or has poor comfort. The specific value of the preset micro-threshold can be set based on actual needs. The preset micro-threshold can be pre-set or determined by the macro-strategy layer based on target navigation information.

[0132] If the micro-scores corresponding to the high-level sub-trajectories are all less than the preset micro-threshold, it means that there is no reasonable and comfortable sub-trajectory among the better sub-trajectories considered by the macro-strategy layer and the meso-strategy layer, and the final target trajectory is often obtained from the high-level sub-trajectory. Therefore, in order to avoid the unreasonableness and discomfort of the target trajectory, in this embodiment, when the micro-scores of the high-level sub-trajectories are all less than the preset micro-threshold, an adjustment instruction is sent to the macro-strategy layer to make the macro-strategy layer re-score.

[0133] The adjustment instruction may specifically include the micro scores corresponding to all sub-trajectories or the micro scores corresponding to high-level sub-trajectories, etc.; after receiving the adjustment instruction, the macro-strategy layer lowers the score of the high-level sub-trajectory, thereby updating the score of the entire sub-trajectory. Since the score of the high-level sub-trajectory is lowered, the original high-level sub-trajectory score in the updated sub-trajectory decreases, and the sorting order of the sub-trajectories based on the intermediate score is also updated. The high-level sub-trajectory corresponding to the updated sub-trajectory changes, so that there can be a high-level sub-trajectory with a micro score greater than the preset micro threshold in the updated high-level sub-trajectory; it can be understood that if the onlooker scores corresponding to the updated high-level sub-trajectory are still less than the preset micro threshold, the adjustment instruction is sent to the macro-strategy layer again until there is a high-level sub-trajectory with a micro score greater than the preset micro threshold.

[0134] It is understandable that an adjustment instruction may also be sent to the meso-strategy layer to update the meso-score of each sub-track, that is, to lower the meso-score corresponding to the high-position sub-track.

[0135] When the meso-strategy layer is combined with the macro-scoring, only adjustment instructions can be sent to the macro-strategy layer, and the macro-score obtained by the macro-strategy layer after lowering the score of the high-level sub-trajectory will be given to the meso-strategy layer. The meso-strategy layer can also reduce the score of the high-level sub-trajectory based on the macro-scoring.

[0136] Micro-scoring such as:

[0137] a micro =(acceleration,steering_rate,β risk )

[0138] Among them, a micro is the microscopic score; acceleration is the score corresponding to acceleration; steering_rate is the score corresponding to the steering angular rate; β risk Scoring is based on preset microscopic thresholds.

[0139] If the preset trajectory standard indicates that the constrained lateral acceleration is less than 3.5 and the jitter parameter jerk is less than 5.0, then when it is detected that the preset trajectory standard is not met, the current state can be recoded as the initial condition and the two-step emergency sampling can be completed within 50ms, and the safe trajectory filter can be started at the same time.

[0140] If there is a high-ranking sub-trajectory whose micro-score reaches the preset micro-threshold, then after superimposing the micro-scores, the high-ranking sub-trajectory whose micro-score reaches the preset micro-threshold in the final multimodal score can be ranked higher, while the high-ranking sub-trajectory whose micro-score is lower than the preset micro-threshold will be ranked lower due to its lower micro-score, so that the micro-score of the high-ranking sub-trajectory with the highest final ranking reaches the preset micro-threshold, that is, the target trajectory can meet the requirements of rationality and comfort.

[0141] The communication between different strategy layers of the hierarchical strategy network can adopt a gated structure. The macro strategy layer can send macro scores to the meso strategy layer and the micro strategy layer in real time, and send preset micro thresholds to the micro strategy layer; the meso strategy layer can send meso scores to the micro strategy layer in real time, and the micro strategy layer can feedback adjustment instructions to the macro strategy layer to achieve closed-loop optimization.

[0142] Multi-agents at the meso-level achieve efficient information flow and decision-making collaboration through graph neural networks, improving safety and game-playing capabilities in multi-vehicle scenarios.

[0143] Furthermore, in a fourth embodiment of the trajectory planning method of the present invention proposed based on the first embodiment of the present invention, step S10 includes the following steps:

[0144] Step S11, obtaining an expected driving route based on the vehicle driving information path planning;

[0145] Step S12, determining a plurality of target points on the expected driving route, and generating a plurality of anchor point trajectories based on the current position of the vehicle and the target points;

[0146] Step S13: for each of the anchor point trajectories, superimposing noise on the anchor point trajectories to obtain multiple noise trajectories corresponding to the anchor point trajectories;

[0147] Step S14: combining the anchor point trajectory and the noise trajectory to obtain the multimodal trajectory.

[0148] To improve the diversity and rationality of trajectory generation, this embodiment introduces a dynamic anchor trajectory mechanism. An anchor trajectory refers to a set of reference trajectories or target points used to guide the trajectory generation process.

[0149] The expected driving route is the driving route planned by the navigation system. If the driving destination has been determined, the navigation system will plan the route to the driving destination based on the current position of the vehicle to obtain the expected driving route.

[0150] The target points are determined on the expected driving route. For example, the current position of the vehicle is used as the starting point. A target point is determined at each preset distance on the expected driving route. When the distance between the target point and the vehicle reaches a distance threshold, multiple target points are obtained. The distance threshold is used to indicate the determination range of the target point. The specific value can be set based on actual needs. For example, when crossing an intersection, the distance threshold can be set to 30 meters so that the target point can cover the driving route contained in the intersection scene.

[0151] After determining the target point, an anchor trajectory is generated with the vehicle's current position as the starting point and the target point as the end point. One anchor trajectory is generated for each target point. Based on the target points and anchor trajectories of the expected driving route, the anchor trajectories can cover all possible driving decisions.

[0152] For each anchor point trajectory, noise is superimposed on the anchor point trajectory to generate a set of perturbed noise trajectories. The specific superposition method can be set based on actual needs, such as Gaussian noise:

[0153]

[0154] Among them, a i is the i-th anchor point trajectory; ∈ is zero mean and variance is σ 2 Gaussian noise; is the obtained noise trajectory.

[0155] In this embodiment, anchor point trajectories are obtained based on the expected driving route, so that the adaptability to the environment and navigation information can be enhanced. At the same time, the expansion of noise trajectories is achieved through noise superposition, which greatly improves the diversity of trajectory generation.

[0156] After obtaining the anchor point trajectory and the noise trajectory, they can also be modeled using the Diffusion Probabilistic Model (DPM). The DPM transforms the complex trajectory distribution into a Gaussian distribution that is easy to model through a gradual noise addition and denoising process. The inverse process is learned through a neural network, thereby achieving high-quality and diverse trajectory generation. Specifically, the anchor point trajectory and the noise trajectory are first noised using the DPM:

[0157]

[0158] Among them, x t is the trajectory obtained by adding noise in the tth step; β t is the noise intensity; I is the unit matrix.

[0159] After multiple steps of noise addition, the trajectory will become close to pure noise. In the generation phase, the diffusion probability model denoises the noisy trajectory and gradually restores the random noise to a trajectory that conforms to the true distribution:

[0160] x t-1 =f θ (x t ,t,F BEV ,C)

[0161] Among them, f θ It is a conditional neural network, and the input includes the noisy trajectory x t , time step t, environmental characteristics F BEV , navigation group information C; x t-1 is the trajectory obtained by denoising in step t. Through multiple iterations, the diffusion probability model finally outputs a high-quality trajectory.

[0162] It should be noted that existing diffusion probability models usually require dozens to hundreds of iterations to complete denoising, which is too inefficient and affects the efficiency of target trajectory generation. Therefore, this embodiment adopts DDIM (Denoising Diffusion Implicit Models) to reduce the number of denoising steps, greatly speeding up the inference speed and meeting the real-time requirements of autonomous driving.

[0163] The diffusion probability model can be configured based on actual needs, such as loss function, pre-training, etc. For example:

[0164] When setting the loss function, you can first pre-train the loss function through the diffusion model, and you can also add multimodal consistency, distributed uncertainty suppression and adaptive regularization terms; the standard diffusion loss L diff for:

[0165]

[0166] Standard diffusion loss is used to improve the denoising and reconstruction capabilities of the model;

[0167] Trajectory regression loss L reg :

[0168]

[0169] indicates the formation trajectory, and τ indicates the predicted trajectory;

[0170] Trajectory regression loss ensures the accuracy of generated trajectories and true trajectories;

[0171] Modality selection loss L modal :

[0172]

[0173] y modal is the score that the model gives to each trajectory, is the corresponding truth score;

[0174] Modality selection loss enhances the distinguishability of multimodal outputs;

[0175] Multimodal consistency loss L consist :

[0176]

[0177] i and j are the subscripts of the two trajectories used for comparison; E is the mathematical expectation; KL is the KL divergence;

[0178] Multimodal consistency loss encourages each modality’s output to maintain diversity rather than randomness within a reasonable range by minimizing the KL divergence between different modal distributions;

[0179] Distributed uncertainty regularization L uncert :

[0180]

[0181] Distributed uncertainty regularization penalizes highly uncertain trajectories, improving model stability and security;

[0182] Total pre-training loss L pretrain :

[0183]

[0184] λ is the weight of the corresponding loss term, which can be adaptively adjusted through meta-learning or cross-validation.

[0185] After pre-training, reinforcement learning is performed on the loss function. During the reinforcement learning phase, in addition to the existing implicit reward optimization, explicit reward gradient guidance, and IQL fine-tuning, a hierarchical adaptive weight and risk perception mechanism is introduced:

[0186] Implicit reward optimization L RL-imp :

[0187]

[0188] π θ Strategies for reinforcement learning;

[0189] Used to directly maximize the expected trajectory reward;

[0190] Explicit Reward Gradient Guidance:

[0191]

[0192] Used to directly guide model parameter updates through reward signals;

[0193] IQL fine-tuning weighting:

[0194]

[0195] is the loss function corresponding to the trajectory τ in the diffusion probability model;

[0196] in:

[0197] w(τ)=exp(α·R(τ))

[0198] w(τ) is the reward weighting factor;

[0199] Hierarchical adaptive weight mechanism:

[0200] The loss term of each layer strategy introduces an adaptive weight β l , where l indicates the policy layer, which is dynamically adjusted through meta-learning:

[0201]

[0202] β l These are the weights corresponding to different layers, namely the macro-strategy layer, the meso-strategy layer, and the micro-strategy layer, which are the parameters learned by the model itself; Loss functions corresponding to different layers;

[0203] Automatically adjust learning objectives at different levels based on scenario complexity and strategy performance.

[0204] Risk-aware regularization adds a risk penalty term to all RL losses to suppress the sampling probability of high-risk trajectories:

[0205]

[0206] Among them, Risk(τ) can be calculated based on indicators such as trajectory uncertainty and distance to obstacles.

[0207] The final total loss of reinforcement learning is:

[0208]

[0209] λ is the weight of the corresponding loss term and can be set based on actual needs.

[0210] The specific training method of the diffusion probability model can be set based on actual needs; for example, a multimodal trajectory dataset is first constructed, and a driving scenario library is constructed through driving data collected from real vehicles. The driving scenario library includes the vehicle state vector, a high-precision BEV semantic map, and multimodal true value trajectories. The vehicle state vector can include position, speed, and heading angle; the high-precision BEV semantic map can include lane lines, traffic signs, and obstacle heat maps; the multimodal true value trajectories can be manually annotated with a variety of reasonable driving behavior patterns.

[0211] Design offline reward calculation function R pre (τ), each trajectory is pre-evaluated via a physical simulator:

[0212] R t =λ1R progress +λ2R safety +λ3R comfort +λ4R rules +λ5R efficiency +λ6R risk

[0213] When pre-training the diffusion probability model, you can set the noise injection strategy and supervised learning optimization.

[0214] The noise injection strategy uses a cosine-scheduled noise diffusion process:

[0215]

[0216] Among them, β t From 10 in 500 steps -4 Increases linearly to 0.02.

[0217] The supervised learning optimization adopts an improved trajectory matching strategy: 1. For each training sample, generate K candidate anchor point trajectories; 2. Calculate the L2 distance between the candidate trajectories and the true value, and select the top-3 trajectories to construct the contrastive learning loss L contrast :

[0218]

[0219] Jointly optimize the noise prediction loss and contrast loss to obtain the total loss function L total :

[0220]

[0221] ∈ is the noise predicted by the model; It is artificially added noise;

[0222] Diffusion model fine-tuning: Designing a reward-guided gradient backpropagation mechanism:

[0223]

[0224] Among them, η uses a dynamic learning rate; θ diff are all parameters of the diffusion probability model:

[0225]

[0226] Reward-guided diffusion sampling adopts a gradient-guided mechanism, injecting reward gradients into the sampling process of the diffusion probability model to form a hybrid update rule:

[0227]

[0228] is the derivative of reward;

[0229] The dynamic guidance coefficient γ t Designed to:

[0230]

[0231] γ base is the basic guidance strength, and the specific value can be set based on actual needs, such as the default 0.15; T is the number of diffusion steps; t is the actual number of steps; the exponential decay term weakens the guidance strength as the denoising process progresses to avoid overfitting in the later stage. For example, a multi-stage sampling strategy is set, with coarse sampling in the early stage, the sampling steps set to 5, the noise level set to high noise, and the guidance strength set to 3; and fine sampling is performed in the later stage, with the sampling steps set to 3, the noise level set to low noise, and the guidance strength set to 0.1.

[0232] To accurately reflect the diversity of autonomous driving trajectories in complex environments, the final multimodal trajectory can also output the probability distribution of the trajectory through a multimodal classification network:

[0233]

[0234] Among them, τ is the sub-trajectory; π m is the weight of the mth sub-trajectory, μ m and Σ m are the mean and covariance respectively.

[0235] Furthermore, the step S14 includes the following steps:

[0236] Step S141, determining the trajectory driving intention corresponding to each of the anchor point trajectories and the noise trajectory;

[0237] Step S142, obtaining a target driving intention corresponding to the expected driving route;

[0238] Step S143, determining the trajectory driving intention and the target driving intention in the anchor point trajectory and the noise trajectory. Figure 1 The corresponding sub-trajectory;

[0239] Step S144: Aggregate the sub-trajectories to obtain the multimodal trajectory.

[0240] Trajectory driving intention is used to indicate the driving tendency of the trajectory, such as the direction of the trajectory, the lane the trajectory is traveling in, etc.

[0241] It can be understood that since the anchor trajectory is obtained based on the target point, and the noise trajectory is obtained by superimposing noise on the anchor trajectory, the specific trajectory driving intentions of different trajectories are not exactly the same. Due to the characteristics of the target point position and the randomness brought by the noise, the driving intention of the trajectory may not necessarily meet the requirements of the expected driving route. When generating the trajectory, the driving intention of the trajectory can be determined, and the trajectories can be grouped based on the trajectory driving intention to obtain navigation grouping information. For example, the trajectory driving intentions of the anchor trajectory and the noise trajectory include turning left, going straight, turning right, etc., and the target driving intention corresponding to the expected driving route is turning left. The trajectory with the trajectory driving intention of turning left meets the requirements of the expected driving route. Therefore, the trajectory driving intention is combined with the target driving intention. Figure 1 The consistent anchor trajectory or noise trajectory is used as a sub-trackline, and subsequent operations are performed based on the obtained sub-trackline.

[0242] Trajectories are filtered to obtain sub-trajectories based on the target driving intention, so that all sub-trajectories in the multimodal trajectory can meet the intention requirements of the expected driving route, thereby reducing the calculation of irrelevant trajectories. Subsequently, the target trajectory is determined only for the sub-trajectories, which improves planning efficiency.

[0243] It should be noted that for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0244] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0245] The present application also provides a trajectory planning device for implementing the above trajectory planning method, the trajectory planning device comprising:

[0246] a first acquisition module, configured to acquire vehicle driving information and generate a multimodal trajectory according to the vehicle driving information, wherein the multimodal trajectory includes a plurality of sub-trajectories;

[0247] a second acquisition module, configured to acquire a hierarchical strategy network model, and input the multimodal trajectory and the vehicle driving information into the hierarchical strategy network model to obtain a multidimensional score corresponding to each sub-trajectory, wherein the hierarchical strategy network model includes multiple strategy layers, each strategy layer scores each sub-trajectory based on different scoring rules combined with the vehicle driving information, and the multidimensional score is determined comprehensively by the scores of each strategy layer;

[0248] A first determination module is configured to determine a target track from the plurality of sub-tracks according to the multi-dimensional scores.

[0249] This trajectory planning device sets up a hierarchical strategy network model consisting of multiple strategy layers, which enables sub-trajectories to be scored based on different scoring dimensions through different strategy layers. This allows the target trajectory to be determined based on the multi-dimensional driving requirements, so that the obtained target trajectory can meet the vehicle driving requirements and adapt to the dynamic driving environment in actual scenarios.

[0250] It should be noted that the first acquisition module in this embodiment can be used to execute step S10 in the embodiment of the present application, the second acquisition module in this embodiment can be used to execute step S20 in the embodiment of the present application, and the first determination module in this embodiment can be used to execute step S30 in the embodiment of the present application.

[0251] Furthermore, the hierarchical strategy network model includes a macro-strategy layer, a meso-strategy layer, and a micro-strategy layer; the second acquisition module includes:

[0252] a first acquiring unit, configured to acquire target navigation information from the vehicle driving information, and score the sub-trajectory according to the target navigation information through the macro-strategy layer to obtain a macro score, wherein the macro score indicates the degree to which the sub-trajectory conforms to the navigation information;

[0253] a second acquisition unit, configured to acquire environmental information from the vehicle driving information, and score the sub-trajectory according to the environmental information through the meso-strategy layer to obtain a meso-score, wherein the meso-score indicates a legality of an interaction between the sub-trajectory and the environmental information;

[0254] a third obtaining unit, configured to obtain a preset trajectory standard, and score the sub-trajectory according to the preset trajectory standard by the micro-strategy layer to obtain a micro-score, wherein the micro-score indicates the degree to which the sub-trajectory meets the preset trajectory standard;

[0255] The first comprehensive unit is used to comprehensively analyze the macro score, the meso score, and the micro score to obtain the multi-dimensional score.

[0256] Furthermore, the meso-strategy layer scores the sub-trajectory according to the environmental information to obtain a meso-score, wherein the meso-score, indicating the legality of the interaction between the sub-trajectory and the environmental information, includes:

[0257] A first acquisition subunit is configured to acquire the predicted trajectory of the surrounding vehicle and characteristic object information from the environmental information;

[0258] A first determining subunit is configured to determine an interactive prediction between the sub-trajectory, the predicted trajectory of the surrounding vehicle, and the feature object;

[0259] The second acquisition subunit is configured to acquire the macro score and determine the meso score based on the interaction prediction and the macro score.

[0260] Furthermore, the micro-strategy layer scores the sub-trajectory according to the preset trajectory standard to obtain a micro-score, wherein the micro-score indicates the degree to which the sub-trajectory meets the preset trajectory standard, including:

[0261] A third acquisition subunit is used to acquire the kinematic standard and the comfort standard in the preset trajectory standard;

[0262] The first scoring sub-unit is used to score each of the sub-trajectories according to the kinematic standard and the comfort standard to obtain the micro score.

[0263] Furthermore, the first determining module includes:

[0264] a first determining unit, configured to obtain an intermediate score corresponding to the sub-trajectory according to the macro score and the meso score;

[0265] a second determining unit, configured to determine a plurality of high-ranking sub-trajectories in the sub-trajectories, wherein the high-ranking sub-trajectories are a preset number of the sub-trajectories with correspondingly higher intermediate scores;

[0266] A first judgment unit is used to judge whether the microscopic scores corresponding to the high-level sub-trajectories all reach a preset microscopic threshold;

[0267] a first sending unit configured to send an adjustment instruction to the macro-strategy layer if the micro-scores corresponding to all the high-order sub-trajectories are less than a preset micro-threshold, so as to cause the macro-strategy layer to reduce the macro-scores corresponding to the high-order sub-trajectories, thereby updating the macro-scores;

[0268] The third determining unit is configured to determine the target trajectory in the high-ranking sub-trajectory according to the multi-dimensional score if there is a microscopic score corresponding to the high-ranking sub-trajectory greater than a preset microscopic threshold.

[0269] Furthermore, the first obtaining includes:

[0270] A first planning unit, configured to obtain an expected driving route based on the vehicle driving information path planning;

[0271] a fourth determining unit, configured to determine a plurality of target points on the expected driving route, and generate a plurality of anchor point trajectories based on the current position of the vehicle and the target points;

[0272] A first execution unit is configured to perform noise superposition on each anchor point trajectory to obtain a plurality of noise trajectories corresponding to the anchor point trajectory;

[0273] The first aggregation unit is configured to aggregate the anchor point trajectory and the noise trajectory to obtain the multimodal trajectory.

[0274] Furthermore, the first set unit includes:

[0275] a second determining subunit, configured to determine a trajectory driving intention corresponding to each of the anchor point trajectories and the noise trajectory;

[0276] A fourth acquisition subunit is configured to acquire a target driving intention corresponding to the expected driving route;

[0277] The fifth determining subunit is configured to determine the trajectory driving intention and the target driving intention in the anchor point trajectory and the noise trajectory. Figure 1 The corresponding sub-trajectory;

[0278] The first aggregation subunit is configured to aggregate the sub-trajectories to obtain the multimodal trajectory.

[0279] Reference Figure 4 In terms of hardware structure, the vehicle may include components such as a communication module 10, a memory 20, and a processor 30. In the vehicle, the processor 30 is connected to the memory 20 and the communication module 10, respectively. The memory 20 stores a computer program, which is simultaneously executed by the processor 30. When the computer program is executed, the steps of the above-mentioned method embodiment are implemented.

[0280] The communication module 10 can be connected to an external communication device via a network. The communication module 10 can receive requests from the external communication device and can also send requests, instructions and information to the external communication device. The external communication device can be another vehicle, a server or an IoT device such as a TV.

[0281] Memory 20 can be used to store software programs and various data. Memory 20 may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function (such as obtaining vehicle driving information). The data storage area may include a database and may store data or information generated based on system usage. Furthermore, memory 20 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0282] Processor 30 is the vehicle's control center, connecting all vehicle components via various interfaces and circuits. By running or executing software programs and / or modules stored in memory 20 and accessing data stored in memory 20, it performs various vehicle functions and processes data, thereby providing overall vehicle monitoring. Processor 30 may include one or more processing units; optionally, processor 30 may integrate an application processor and a modem processor. The application processor primarily handles the operating system, user interface, and application programs, while the modem processor primarily handles wireless communications. It is understood that the modem processor may not be integrated into processor 30.

[0283] although Figure 4Although not shown, the above vehicle may further include a circuit control module, which is used to connect to the power supply to ensure the normal operation of other components. Figure 4 The vehicle structure shown in the figure does not constitute a limitation to the vehicle, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0284] The present invention also provides a computer-readable storage medium on which a computer program is stored. The computer-readable storage medium may be Figure 4 The memory 20 in the vehicle may also be at least one of a ROM (Read-Only Memory) / RAM (Random Access Memory), a magnetic disk, and an optical disk. The computer-readable storage medium includes a number of instructions for enabling a terminal device with a processor (which may be a television, a car, a mobile phone, a computer, a server, a terminal, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

[0285] In the present invention, the terms "first", "second", "third", "fourth" and "fifth" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0286] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0287] Although the embodiments of the present invention have been shown and described above, the scope of protection of the present invention is not limited thereto. It should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Persons skilled in the art may make changes, modifications, and substitutions to the above embodiments within the scope of the present invention, and such changes, modifications, and substitutions should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A trajectory planning method, characterized in that: The trajectory planning method comprises: Acquiring vehicle driving information, and generating a multimodal trajectory based on the vehicle driving information, wherein the multimodal trajectory includes a plurality of sub-trajectories; Obtaining a hierarchical policy network model, and inputting the multimodal trajectory and the vehicle driving information into the hierarchical policy network model to obtain a multidimensional score corresponding to each sub-trajectory, wherein the hierarchical policy network model includes multiple policy layers, each policy layer scores each sub-trajectory based on different scoring rules combined with the vehicle driving information, and the multidimensional score is comprehensively determined by the scores of each policy layer; A target track is determined from the plurality of sub-tracks according to the multi-dimensional scores.

2. The trajectory planning method according to claim 1, wherein: The hierarchical strategy network model includes a macro-strategy layer, a meso-strategy layer, and a micro-strategy layer; inputting the multimodal trajectory and the vehicle driving information into the hierarchical strategy network model to obtain a multi-dimensional score corresponding to each sub-trajectory includes: Obtaining target navigation information from the vehicle driving information, and scoring the sub-trajectory according to the target navigation information by the macro-strategy layer to obtain a macro score, wherein the macro score indicates the degree to which the sub-trajectory conforms to the navigation information; Obtaining environmental information from the vehicle driving information, and scoring the sub-trajectory according to the environmental information through the meso-strategy layer to obtain a meso-score, wherein the meso-score indicates the legality of the interaction between the sub-trajectory and the environmental information; Obtaining a preset trajectory standard, and scoring the sub-trajectory according to the preset trajectory standard by the micro-strategy layer to obtain a micro-score, wherein the micro-score indicates the degree to which the sub-trajectory meets the preset trajectory standard; The multi-dimensional score is obtained by integrating the macro score, the meso score and the micro score.

3. The trajectory planning method according to claim 2, wherein: The meso-level score is obtained by scoring the sub-trajectory according to the environmental information through the meso-level strategy layer, wherein the meso-level score indicates the legality of the interaction between the sub-trajectory and the environmental information, including: Obtaining the predicted trajectory of the surrounding vehicle and characteristic object information in the environmental information; Determining interactive predictions between the sub-trajectory, the surrounding vehicle predicted trajectory, and the feature object; The macro score is obtained, and the meso score is determined based on the interactive prediction and the macro score.

4. The trajectory planning method according to claim 2, wherein: Scoring the sub-trajectory according to the preset trajectory standard by the micro-strategy layer to obtain a micro-score, wherein the micro-score indicates the degree to which the sub-trajectory meets the preset trajectory standard, includes: Obtaining a kinematic standard and a comfort standard in the preset trajectory standard; Each of the sub-trajectories is scored according to the kinematic standard and the comfort standard to obtain the micro-score.

5. The trajectory planning method according to claim 2, wherein: Determining a target track from the plurality of sub-tracks according to the multi-dimensional scores includes: Obtaining an intermediate score corresponding to the sub-track according to the macro score and the meso score; Determining a plurality of high-ranking sub-trajectories in the sub-trajectories, wherein the high-ranking sub-trajectories are a preset number of the sub-trajectories with correspondingly higher intermediate scores; Determine whether the microscopic scores corresponding to the high-order sub-trajectories all reach a preset microscopic threshold; If the micro scores corresponding to all the high-level sub-trajectories are less than the preset micro threshold, an adjustment instruction is sent to the macro-strategy layer to cause the macro-strategy layer to reduce the macro scores corresponding to the high-level sub-trajectories to update the macro scores; If there is a microscopic score corresponding to the high-order sub-trajectory that is greater than a preset microscopic threshold, the target trajectory is determined in the high-order sub-trajectory according to the multi-dimensional score.

6. The trajectory planning method according to claim 1, wherein: Generating a multimodal trajectory according to the vehicle driving information includes: Obtaining an expected driving route based on the vehicle driving information path planning; Determining a plurality of target points on the expected driving route, and generating a plurality of anchor point trajectories based on the current position of the vehicle and the target points; For each of the anchor point trajectories, performing noise superposition on the anchor point trajectories to obtain multiple noise trajectories corresponding to the anchor point trajectories; The anchor point trajectory and the noise trajectory are combined to obtain the multimodal trajectory.

7. The trajectory planning method according to claim 6, wherein: The combining the anchor point trajectory and the noise trajectory to obtain the multimodal trajectory includes: determining a trajectory driving intention corresponding to each of the anchor point trajectories and the noise trajectory; Obtaining a target driving intention corresponding to the expected driving route; Determining a sub-trajectory in the anchor point trajectory and the noise trajectory, wherein the trajectory driving intention is consistent with the target driving intention; The sub-trajectories are aggregated to obtain the multimodal trajectory.

8. A trajectory planning device, characterized in that: The trajectory planning device comprises: a first acquisition module, configured to acquire vehicle driving information and generate a multimodal trajectory according to the vehicle driving information, wherein the multimodal trajectory includes a plurality of sub-trajectories; a second acquisition module, configured to acquire a hierarchical strategy network model, and input the multimodal trajectory and the vehicle driving information into the hierarchical strategy network model to obtain a multidimensional score corresponding to each sub-trajectory, wherein the hierarchical strategy network model includes multiple strategy layers, each strategy layer scores each sub-trajectory based on different scoring rules combined with the vehicle driving information, and the multidimensional score is determined comprehensively by the scores of each strategy layer; A first determination module is configured to determine a target track from the plurality of sub-tracks according to the multi-dimensional scores.

9. A vehicle, characterized in that: The vehicle includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the trajectory planning method according to any one of claims 1 to 7 when executed by the processor.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the trajectory planning method according to any one of claims 1 to 7.

Images