Strategy planning method, device and equipment based on self-adaptive trajectory fine tuning, medium and product
Through self-attention transformer network and local fine-tuning method, the information interaction between bicycle candidate trajectories and dynamic environments is enhanced, and the problem of insufficient accuracy and adaptability in the construction process of trajectory tree in the existing technology is solved, and more efficient strategic planning is achieved.
Patent Information
- Application Number
- CN202510426136.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-07
- Publication Date
- 2025-08-15
AI Technical Summary
Existing tree-based strategy planning method In the process of trajectory tree construction, the generation of bicycle candidate trajectories lacks information interaction with the dynamic environment, resulting in limited candidate trajectory accuracy and scene adaptability.
Through the self-attention transformer network, the information interaction between the self-agent and surrounding agents and road map elements is achieved, multimodal candidate trajectories are generated, and the trajectory is adjusted using local fine-tuning methods to build a trajectory tree and a scene tree, and select the optimal strategy in combination with the Markov decision-making process.
It improves the accuracy and scenario adaptability of the bicycle candidate trajectory, and improves the accuracy and effectiveness of strategy planning.
Smart Images
Figure BDA0005346776520000082 
Figure HDA0005346776530000011 
Figure HDA0005346776530000021
Abstract
Description
Technical Field
[0001] The present application relates to the fields of control principles, automation technology, and data processing, and in particular to a strategy planning method, device, equipment, medium, and product based on adaptive trajectory fine-tuning. Background Art
[0002] In dynamic interactive environments, achieving safe, comfortable, and socially acceptable policy planning is crucial for the large-scale deployment of autonomous vehicles. Existing tree-based policy planning methods transform the continuous-space motion planning problem into a finite-time Markov decision process (MDP) by constructing trajectory trees and scenario trees, providing a theoretical foundation and implementation path for policy planning. A typical process includes:
[0003] (1) Generate multimodal candidate trajectories of the ego vehicle through sampling methods and construct a trajectory tree;
[0004] (2) Predict the future trajectories of surrounding agents based on the candidate trajectories of the ego vehicle, and then construct a scenario tree;
[0005] (3) Construct an MDP on the trajectory tree and the scenario tree, and select the optimal strategy through a dynamic process.
[0006] While such methods have advanced the development of policy planning systems, the generation of candidate ego-vehicle trajectories lacks information interaction with the dynamic environment during trajectory tree construction, limiting the accuracy and adaptability of candidate trajectories to specific scenarios, thus impacting the overall policy planning performance. Therefore, it is urgent to study the information interaction mechanism between candidate ego-vehicle trajectories and the dynamic environment during trajectory tree construction to improve the accuracy and adaptability of policy planning. Summary of the Invention
[0007] The purpose of this application is to provide a strategy planning method, device, equipment, medium and product based on adaptive trajectory fine-tuning, which solves the problem that the generation of candidate trajectories of the vehicle lacks information interaction with the dynamic environment, and the accuracy and adaptability of the candidate trajectories to the scene are limited.
[0008] To achieve the above objectives, this application provides the following solutions:
[0009] In a first aspect, the present application provides a strategy planning method based on adaptive trajectory fine-tuning, comprising:
[0010] Obtain input information for strategy planning; the input information includes historical trajectory information of the self-agent, current state information of the self-agent, historical trajectory information of surrounding agents, road map element information and lane map information.
[0011] Based on the self-attention transformer network, information interaction between the self-agent and surrounding agents, as well as between the self-agent and road map elements, is realized to obtain an environmental context containing dynamic interaction information.
[0012] Based on the environmental context and according to the current state information of the self-agent, a sampling method is used to generate a multimodal candidate trajectory of the self-agent; the multimodal candidate trajectory is a plurality of candidate trajectories.
[0013] Anchor points are sampled along each candidate trajectory of the ego-agent, and environmental context information around the anchor points is extracted.
[0014] Based on the environmental context information around the anchor point, each candidate trajectory of the ego agent is adjusted using a local fine-tuning method, and the adjusted candidate trajectories are used to construct a trajectory tree of the ego agent.
[0015] Based on the adjusted candidate trajectories in the trajectory tree, a conditional prediction method is used to predict the future trajectory of the surrounding intelligent agent to obtain a prediction result.
[0016] Based on the prediction results, a scene tree is constructed.
[0017] A Markov decision process is constructed on the trajectory tree and the scene tree, and a cost function is used to evaluate the cost of each adjusted candidate trajectory in the trajectory tree, and the trajectory with the lowest cost is selected as the optimal strategy.
[0018] In a second aspect, the present application provides a strategy planning device based on adaptive trajectory fine-tuning, comprising:
[0019] The input information acquisition module is used to obtain input information for strategy planning; the input information includes the historical trajectory information of the self-agent, the current state information of the self-agent, the historical trajectory information of the surrounding agents, road map element information and lane map information.
[0020] The environmental context determination module is used to realize information interaction between the self-agent and surrounding agents, as well as between the self-agent and road map elements based on the self-attention transformer network, and obtain the environmental context containing dynamic interaction information.
[0021] The multimodal candidate trajectory determination module of the self-agent is used to generate the multimodal candidate trajectory of the self-agent based on the environmental context and the current state information of the self-agent using a sampling method; the multimodal candidate trajectory is a plurality of candidate trajectories.
[0022] The anchor point selection and environmental context information extraction module is used to sample anchor points along each candidate trajectory of the self-agent and extract environmental context information around the anchor points.
[0023] A trajectory tree construction module is used to adjust each candidate trajectory of the self-agent using a local fine-tuning method based on the environmental context information around the anchor point, and to construct a trajectory tree of the self-agent using the adjusted candidate trajectory.
[0024] The surrounding agent trajectory prediction module is used to predict the future trajectory of the surrounding agent based on the adjusted candidate trajectory in the trajectory tree using a conditional prediction method to obtain a prediction result.
[0025] The scene tree construction module is used to construct a scene tree based on the prediction result.
[0026] The optimal strategy determination module is used to construct a Markov decision process on the trajectory tree and the scene tree, and use a cost function to evaluate the cost of each adjusted candidate trajectory in the trajectory tree, and select the trajectory with the lowest cost as the optimal strategy.
[0027] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any one of the above-mentioned strategy planning methods based on adaptive trajectory fine-tuning.
[0028] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the above-mentioned strategy planning methods based on adaptive trajectory fine-tuning.
[0029] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned strategy planning methods based on adaptive trajectory fine-tuning.
[0030] According to the specific embodiments provided in this application, this application has the following technical effects:
[0031] The present application provides a strategy planning method, apparatus, device, medium and product based on adaptive trajectory fine-tuning, which obtains input information for strategy planning; realizes information interaction between the self-agent and surrounding agents, as well as between the self-agent and road map elements based on the self-attention transformer network, and obtains an environmental context containing dynamic interaction information; based on the current state of the self-agent, generates the multimodal candidate trajectory of the self-agent using a sampling method; samples anchor points along each candidate trajectory of the self-agent, and dynamically adjusts the search radius around the anchor point according to the speed of the agent and the number of iterations of trajectory optimization, so as to fully capture the environment around the anchor point. Environmental context information; based on the environmental context information around the anchor point, each candidate trajectory of the self-agent is adjusted using local fine-tuning and iterative optimization methods, thereby improving the accuracy and scene adaptability of the candidate trajectory of the self-car, and using the adjusted candidate trajectory to construct a trajectory tree of the self-agent; based on each candidate trajectory in the trajectory tree, a conditional prediction method is used to predict the future trajectory of the surrounding agents to obtain a prediction result; based on the prediction result, a scene tree is constructed; a Markov decision process is constructed on the trajectory tree and the scene tree, and a cost function is used to evaluate the cost of each candidate trajectory in the trajectory tree, and the trajectory with the lowest cost is selected as the optimal strategy. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.
[0033] Figure 1 Schematic diagram of the process of the strategy planning method based on adaptive trajectory fine-tuning in an embodiment of the present application.
[0034] Figure 2 Schematic diagram of the working process of the strategy planning method based on adaptive trajectory fine-tuning in an embodiment of the present application.
[0035] Figure 3 Schematic diagram of the principle of the strategy planning method based on adaptive trajectory fine-tuning in an embodiment of the present application.
[0036] Figure 4 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0037] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.
[0038] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0039] like Figure 1 As shown, a strategy planning method based on adaptive trajectory fine-tuning includes input information; the input information includes: historical trajectory information of the self-vehicle agent, current state information of the self-vehicle agent, historical trajectory information of N surrounding agents, W surrounding road map element information and lane map information; the historical trajectory information of the self-vehicle agent and the surrounding agents is encoded using an LSTM network, and the road map element information is encoded using an MLP network; the encoding results are spliced into feature vectors, and a self-attention transformer is used to realize information interaction between the self-vehicle agent and the surrounding agents, and between the self-vehicle agent and the road map elements, to generate an environmental context containing dynamic interaction information; based on the current state of the self-vehicle agent, a set of position information at the target moment is sampled along the center line of the lane map; a spline curve from the current position to the target position is fitted using a third-order polynomial to generate a multimodal candidate trajectory for the self-vehicle agent; Each candidate trajectory of the ego-agent is divided into D segments, and the endpoints of each segment are selected as anchor points; the search radius around the anchor points is dynamically adjusted based on the agent's speed v and the number of trajectory optimization iterations k, and the environmental context information around each anchor point is extracted; the extracted environmental context information is converted from a global coordinate system centered on the ego-agent to a local coordinate system centered on the anchor point; each trajectory segment on the candidate trajectory is adjusted using a local fine-tuning method; after local fine-tuning of all trajectory segments, the candidate trajectory is optimized once; through k iterative optimizations, the accuracy and adaptability of the candidate trajectory of the ego-agent to the scene are improved; the adjusted candidate trajectories are used to construct a trajectory tree of the ego-agent; using a conditional prediction method, the future trajectories of the surrounding agents are predicted based on the candidate trajectory of the ego-agent, and a scene tree is constructed based on the prediction results; a Markov decision process (MDP) is constructed on the trajectory tree and the scene tree, and a cost function is used to evaluate the cost of each candidate trajectory of the ego-agent in the trajectory tree, and the trajectory with the lowest cost is selected as the optimal strategy.
[0040] like Figure 2As shown, the embodiment of the present application provides a strategy planning method based on adaptive trajectory fine-tuning, which specifically includes:
[0041] S1: Obtain input information for strategy planning; the input information includes historical trajectory information of the self-agent, current state information of the self-agent, historical trajectory information of surrounding agents, road map element information, and lane map information.
[0042] S2: Based on the self-attention transformer network, information interaction is realized between the self-agent and surrounding agents, as well as between the self-agent and road map elements, to obtain an environmental context containing dynamic interaction information.
[0043] S3: Based on the environmental context and according to the current state information of the ego agent, a sampling method is used to generate a multimodal candidate trajectory of the ego agent; the multimodal candidate trajectory is a plurality of candidate trajectories.
[0044] S4: Sample anchor points along each candidate trajectory of the ego-agent and extract environmental context information around the anchor points.
[0045] S5: Based on the environmental context information around the anchor point, each candidate trajectory of the ego agent is adjusted using a local fine-tuning method, and the adjusted candidate trajectories are used to construct a trajectory tree of the ego agent.
[0046] S6: Based on the adjusted candidate trajectories in the trajectory tree, a conditional prediction method is used to predict the future trajectories of the surrounding intelligent agents to obtain a prediction result.
[0047] S7: Construct a scenario tree based on the prediction results.
[0048] S8: Constructing a Markov decision process on the trajectory tree and the scene tree, and using a cost function to evaluate the cost of each adjusted candidate trajectory in the trajectory tree, and selecting the trajectory with the lowest cost as the optimal strategy.
[0049] Further, in an exemplary embodiment, step S2 may be replaced by the following steps.
[0050] S201: Encode the historical trajectory information of the self-agent and the surrounding agents according to the LSTM network, and encode the information of the road map elements according to the MLP network to obtain an encoding result.
[0051] S202: Based on the encoding result, information interaction between the self-agent and the surrounding agents, and between the self-agent and the road map elements is realized according to the self-attention transformer network to obtain an environmental context containing dynamic interaction information.
[0052] The encoding results are concatenated into a vector containing [1+N+W]-dimensional features, and information interaction between the self-agent and surrounding agents, as well as between the self-agent and road map elements, is realized through the self-attention transformer; where N is the number of agents around the self-vehicle; W is the number of map elements around the self-vehicle.
[0053] Further, in an exemplary embodiment, step S3 may be replaced by the following steps.
[0054] S301: Based on the current state of the ego agent, sample the position information of the target moment along the center line of the lane map; the position information includes the current position and the target position.
[0055] Based on the current state of the vehicle, the target time T is sampled along the center line of the lane map j The location information includes the current location and the target location.
[0056] S302: Using a third-order polynomial to fit a spline curve from the current position to the target position, to obtain a multimodal candidate trajectory of the ego agent.
[0057] As a specific embodiment, the current state s(0) = [x0, y0, v0, h0], the target state Use two independent third-order polynomials x(t) and y(t) to parameterize the x and y coordinates so that a spline curve connects the two points and satisfies the following boundary conditions:
[0058]
[0059] Using the above boundary conditions, the coefficients of two third-order polynomials can be uniquely determined to generate the candidate trajectory of the vehicle; where x0 is the horizontal coordinate of the trajectory at the current moment; y0 is the vertical coordinate of the trajectory at the current moment; v0 is the speed at the current moment; h0 is the heading angle at the current moment; x Tj is the horizontal coordinate of the trajectory at time Tj; y Tj is the vertical coordinate of the trajectory at time Tj; v Tj is the speed at time Tj; h Tj is the heading angle at time Tj; Tj is time.
[0060] Further, in an exemplary embodiment, step S4 may be replaced by the following steps.
[0061] S401: Divide each candidate trajectory in the trajectory tree into D segments, and select the endpoint of each trajectory segment as an anchor point; D is the number of segments.
[0062] Each candidate trajectory in the trajectory tree is divided into D = α·Tj segments, where α represents the scale factor, T j For the moment.
[0063] S402: Based on the speed of the ego agent and the number of iterations of the trajectory fine-tuning, adjust the search radius around the anchor point, and extract the environmental context information around each anchor point within the search radius.
[0064] S403: Use the formula Rad(v,k)=f1(v)·f2(k) to adjust the retrieval radius of the anchor point; wherein, Rad(v,k) is the retrieval radius of the anchor point when the speed of the ego agent is v and the number of iterations of trajectory fine-tuning is k; v is the speed of the ego agent; k is the number of iterations of trajectory fine-tuning; f1() is a monotonically increasing function, f1() is used to expand the retrieval radius as the speed of the ego agent increases; f2() is a monotonically decreasing function, f2() is used to capture local fine-grained features around the anchor point in the later stage of the optimization process.
[0065] Further, in an exemplary embodiment, step S5 may be replaced by the following steps.
[0066] S501: Convert the environmental context information from a global coordinate system centered on the ego agent to a local coordinate system centered on an anchor point.
[0067] S502: Based on the local coordinate system, adjust each trajectory segment on the candidate trajectory using a local fine-tuning method to obtain an adjusted candidate trajectory.
[0068] In each local fine-tuning step, the cross-attention mechanism is used to fuse the features of the candidate ego-vehicle trajectory with the contextual information around the corresponding anchor point. Based on the fused features, the offset of the path point on the trajectory segment is predicted and added to the original trajectory segment to achieve local trajectory fine-tuning. After local fine-tuning of all trajectory segments, the candidate trajectory is optimized once. Through k-times of iterative optimization, the accuracy and adaptability of the candidate trajectory to the scene are improved.
[0069] S503: Constructing a trajectory tree of the ego agent using the adjusted candidate trajectory.
[0070] The conditional prediction method in the field of trajectory prediction is to predict the future trajectory of surrounding agents based on the trajectory of the self-agent as a condition (or prior information).
[0071] An MDP is constructed on the trajectory tree and scenario tree. A cost function is used to evaluate the cost of each candidate trajectory in the trajectory tree, and the trajectory with the lowest cost is selected as the optimal policy. The cost function includes cost items such as lane keeping, collision avoidance, ride comfort, and goal achievement.
[0072] like Figure 3 As shown, the embodiment of the present application provides a strategy planning device based on adaptive trajectory fine-tuning, specifically including:
[0073] The input information acquisition module is used to obtain input information for strategy planning; the input information includes the historical trajectory information of the self-agent, the current state information of the self-agent, the historical trajectory information of the surrounding agents, road map element information and lane map information.
[0074] The environmental context determination module is used to realize information interaction between the self-agent and surrounding agents, as well as between the self-agent and road map elements based on the self-attention transformer network, and obtain the environmental context containing dynamic interaction information.
[0075] The multimodal candidate trajectory determination module of the self-agent is used to generate the multimodal candidate trajectory of the self-agent based on the environmental context and the current state information of the self-agent using a sampling method; the multimodal candidate trajectory is a plurality of candidate trajectories.
[0076] The anchor point selection and environmental context information extraction module is used to sample anchor points along each candidate trajectory of the self-agent and extract environmental context information around the anchor points.
[0077] A trajectory tree construction module is used to adjust each candidate trajectory of the self-agent using a local fine-tuning method based on the environmental context information around the anchor point, and to construct a trajectory tree of the self-agent using the adjusted candidate trajectory.
[0078] The surrounding agent trajectory prediction module is used to predict the future trajectory of the surrounding agent based on the adjusted candidate trajectory in the trajectory tree using a conditional prediction method to obtain a prediction result.
[0079] The scene tree construction module is used to construct a scene tree based on the prediction result.
[0080] The optimal strategy determination module is used to construct a Markov decision process on the trajectory tree and the scene tree, and use a cost function to evaluate the cost of each adjusted candidate trajectory in the trajectory tree, and select the trajectory with the lowest cost as the optimal strategy.
[0081] The present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any one of the above-described adaptive trajectory fine-tuning strategy planning methods.
[0082] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store video tag processing data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a video tag processing method is implemented.
[0083] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements any of the above-mentioned strategy planning methods based on adaptive trajectory fine-tuning.
[0084] An embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned strategy planning methods based on adaptive trajectory fine-tuning.
[0085] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0086] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A strategy planning method based on adaptive trajectory fine-tuning, characterized in that: The strategy planning method based on adaptive trajectory fine-tuning includes: Obtaining input information for strategy planning; the input information includes historical trajectory information of the self-agent, current state information of the self-agent, historical trajectory information of surrounding agents, road map element information, and lane map information; Based on the self-attention transformer network, information interaction between the self-agent and surrounding agents, as well as between the self-agent and road map elements, is achieved to obtain an environmental context containing dynamic interaction information; Based on the environmental context and according to the current state information of the self-agent, a sampling method is used to generate a multimodal candidate trajectory of the self-agent; the multimodal candidate trajectory is a plurality of candidate trajectories; Sampling anchor points along each candidate trajectory of the ego-agent and extracting environmental context information around the anchor points; Based on the environmental context information around the anchor point, each candidate trajectory of the ego agent is adjusted using a local fine-tuning method, and the adjusted candidate trajectories are used to construct a trajectory tree of the ego agent; Based on the adjusted candidate trajectories in the trajectory tree, using a conditional prediction method to predict the future trajectory of the surrounding intelligent agent to obtain a prediction result; Based on the prediction results, construct a scenario tree; A Markov decision process is constructed on the trajectory tree and the scene tree, and a cost function is used to evaluate the cost of each adjusted candidate trajectory in the trajectory tree, and the trajectory with the lowest cost is selected as the optimal strategy.
2. The strategy planning method based on adaptive trajectory fine-tuning according to claim 1 is characterized in that: Based on the self-attention transformer network, information interaction is achieved between the self-agent and surrounding agents, as well as between the self-agent and road map elements, to obtain an environmental context containing dynamic interaction information, specifically including: Encoding the historical trajectory information of the self-agent and the surrounding agents according to the LSTM network, and encoding the information of the road map elements according to the MLP network to obtain an encoding result; Based on the encoding results, information interaction between the self-agent and the surrounding agents, and between the self-agent and the road map elements is realized according to the self-attention transformer network to obtain an environmental context containing dynamic interaction information.
3. The strategy planning method based on adaptive trajectory fine-tuning according to claim 1, characterized in that: Based on the environmental context and according to the current state information of the ego agent, a sampling method is used to generate a multimodal candidate trajectory of the ego agent, specifically including: Based on the current state of the self-agent, sampling the position information of the target moment along the center line of the lane map; the position information includes the current position and the target position; A spline curve from the current position to the target position is fitted using a third-order polynomial to obtain a multimodal candidate trajectory of the ego agent.
4. The strategy planning method based on adaptive trajectory fine-tuning according to claim 1, characterized in that: Sampling anchor points along each candidate trajectory of the ego-agent and extracting the environmental context information around the anchor points, specifically including: Divide each candidate trajectory in the trajectory tree into D segments, and select the endpoint of each trajectory segment as the anchor point; D is the number of segments; Based on the speed of the ego agent and the number of iterations of the trajectory fine-tuning, the search radius around the anchor point is adjusted, and within the search radius, the environmental context information around each of the anchor points is extracted.
5. The strategy planning method based on adaptive trajectory fine-tuning according to claim 4 is characterized in that: Adjusting the search radius around the anchor point based on the speed of the ego agent and the number of iterations of the trajectory fine-tuning, specifically including: The retrieval radius of the anchor point is adjusted using the formula Rad(v,k)=f1(v)·f2(k); wherein Rad(v,k) is the retrieval radius of the anchor point when the ego agent speed is v and the number of trajectory fine-tuning iterations is k; v is the ego agent speed; k is the number of trajectory fine-tuning iterations; f1() is a monotonically increasing function, which is used to expand the retrieval radius as the ego agent speed increases; f2() is a monotonically decreasing function, which is used to capture local fine-grained features around the anchor point in the later stage of the optimization process.
6. The strategy planning method based on adaptive trajectory fine-tuning according to claim 1, characterized in that: Based on the environmental context information around the anchor point, each candidate trajectory of the ego agent is adjusted using a local fine-tuning method, and the adjusted candidate trajectories are used to construct a trajectory tree of the ego agent, specifically including: Converting the environmental context information from a global coordinate system centered on the ego-agent to a local coordinate system centered on the anchor point; Based on the local coordinate system, adjusting each trajectory segment on the candidate trajectory using a local fine-tuning method to obtain an adjusted candidate trajectory; The trajectory tree of the ego agent is constructed using the adjusted candidate trajectories.
7. A strategy planning device based on adaptive trajectory fine-tuning, characterized in that: The strategy planning device for adaptive trajectory fine-tuning includes: An input information acquisition module is used to obtain input information for strategy planning; the input information includes historical trajectory information of the self-agent, current state information of the self-agent, historical trajectory information of surrounding agents, road map element information, and lane map information; The environmental context determination module is used to realize information exchange between the self-agent and surrounding agents, as well as between the self-agent and road map elements based on the self-attention transformer network, and obtain the environmental context containing dynamic interaction information; a multimodal candidate trajectory determination module for an ego agent, configured to generate a multimodal candidate trajectory for the ego agent based on the environmental context and according to the current state information of the ego agent using a sampling method; the multimodal candidate trajectory is a plurality of candidate trajectories; An anchor point selection and environmental context information extraction module, for sampling anchor points along each candidate trajectory of the ego agent and extracting environmental context information around the anchor points; a trajectory tree construction module, configured to adjust each candidate trajectory of the ego agent using a local fine-tuning method based on the environmental context information around the anchor point, and construct a trajectory tree of the ego agent using the adjusted candidate trajectories; A surrounding agent trajectory prediction module is configured to predict the future trajectory of the surrounding agent using a conditional prediction method based on the adjusted candidate trajectory in the trajectory tree to obtain a prediction result; A scene tree construction module, configured to construct a scene tree based on the prediction result; The optimal strategy determination module is used to construct a Markov decision process on the trajectory tree and the scene tree, and use a cost function to evaluate the cost of each adjusted candidate trajectory in the trajectory tree, and select the trajectory with the lowest cost as the optimal strategy.
8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the strategy planning method based on adaptive trajectory fine-tuning according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the strategy planning method based on adaptive trajectory fine-tuning according to any one of claims 1 to 6 is implemented.
10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the strategy planning method based on adaptive trajectory fine-tuning according to any one of claims 1 to 6 is implemented.