Vehicle driving planning method and system
By combining scene cognition networks and policy memory, a decision-oriented scene summary descriptor is generated, which activates the corresponding decision modules and integrates historical strategies to generate the final execution strategy. This solves the problems of low efficiency and poor adaptability of existing vehicle driving planning systems and achieves efficient and reliable vehicle driving decisions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUXI HUAIJIN HOLDYU TECHNOLOGY CO LTD
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-21
AI Technical Summary
Existing vehicle driving planning systems are inefficient, error-prone, and poorly adaptable, failing to effectively utilize sensor data to generate efficient and safe driving trajectories and behavioral decisions.
By acquiring vehicle environmental perception data, primary environmental features are extracted and input into the scene cognition network to generate a decision-oriented scene summary descriptor. The hierarchical policy network is activated and similar historical policy memory fragments are retrieved. The final execution policy is generated through fusion via the meta-policy network, and the policy memory bank is updated in real time to achieve closed-loop accumulation.
It significantly improves the vehicle's decision-making speed and reliability in complex road conditions, enhances its adaptive capabilities, and ensures the safety and efficiency of vehicle operation.
Smart Images
Figure CN121898458A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle management technology, and in particular to a vehicle driving planning method and system. Background Technology
[0002] Vehicle driving planning is a key core technology in the fields of intelligent driving and artificial intelligence. It aims to enable vehicles to autonomously generate safe, efficient, and comfortable driving trajectories and behavioral decisions based on their perception and understanding of the environment. Deep learning and reinforcement learning, representing artificial intelligence technologies, are widely used in this field. Imitation learning generates strategies by mimicking human driving behavior data, with end-to-end models directly mapping perceptual inputs to control commands. Reinforcement learning, on the other hand, optimizes long-term cumulative rewards through trial and error in virtual or real environments, allowing vehicles to learn better strategies in interactive game scenarios.
[0003] Existing systems typically first perceive the surrounding environment through sensors such as cameras and radar, generating a general report similar to a checklist. For example, if there is a car 10 meters to the left front at a speed of 50 kilometers per hour, another independent decision-making module needs to analyze this lengthy checklist like solving a puzzle to determine which information is most important for the current driving situation, and thus decide whether to accelerate, decelerate, or turn. This model has two drawbacks: inefficiency and error-proneness, and poor adaptability. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, the present invention provides a vehicle driving planning method to solve the problems of low efficiency, error-proneness, and poor adaptability of the prior art.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a vehicle driving planning method, comprising, S1: Acquire vehicle environmental perception data, process the environmental perception data, and extract primary environmental features; S2: Input the primary environmental features into the scene cognition network to generate a decision-oriented scene summary descriptor, which is used to quantitatively characterize the type and degree of challenge that the current driving scene poses to decision-making; S3: Based on the scene summary descriptor, activate the corresponding upper-level decision module in the hierarchical policy network, and simultaneously retrieve historical policy memory fragments similar to the scene summary descriptor in the policy memory bank; S4: Based on the activated upper-level decision-making module, an endogenous basic policy is generated, and based on the retrieved historical policy memory fragments, the endogenous basic policy and the historical policy kernel are fused and arbitrated through the meta-policy network to generate the final execution policy. S5: Execute the final execution strategy to control vehicle movement and collect the performance indicators of strategy execution in real time; S6: Update the strategy memory based on the scenario summary descriptor, the final execution strategy, and the performance indicators to complete the closed-loop accumulation of strategy knowledge.
[0007] As a preferred embodiment of the vehicle driving planning method of the present invention, in step S2, the primary environmental features are standardized and vector-concatenated to construct a global feature vector, which is then input in parallel into multiple attention mechanism calculation modules in the scene cognition network. Each attention mechanism calculation module corresponds to a preset decision challenge dimension. In each attention mechanism calculation module, the correlation weight between the global feature vector and the decision challenge dimension corresponding to the attention mechanism calculation module is analyzed. The correlation weight represents the degree of influence of elements related to that dimension in the current environment on driving decisions. The correlation weight output by each attention mechanism calculation module is normalized to generate a quantitative scoring sub-vector corresponding to the decision challenge dimension. All quantitative scoring sub-vectors of the decision challenge dimensions are combined to generate the final scene summary descriptor.
[0008] As a preferred embodiment of the vehicle driving planning method described in this invention, the relevance weight analysis process in the relevance weight of the global feature vector and the decision challenge dimension corresponding to the attention mechanism calculation module includes: generating a query vector from the global feature vector through linear transformation; generating a key vector from the feature representation of the decision challenge dimension through linear transformation; performing a dot product operation between the query vector and the key vector and then processing it through a nonlinear activation function to obtain the original relevance score of the decision challenge dimension; and normalizing the original relevance score to generate the relevance weight.
[0009] As a preferred embodiment of the vehicle driving planning method of the present invention, wherein: in step S3, retrieving historical strategy memory fragments in the strategy memory bank based on scene summary descriptors includes the following sub-steps: obtaining multiple historical strategy memory fragments stored in the strategy memory bank, each historical strategy memory fragment containing a historical scene summary descriptor; analyzing the similarity between the current scene summary descriptor and each historical scene summary descriptor using the cosine similarity method; sorting the historical strategy memory fragments in descending order of similarity; and selecting the top N historical strategy memory fragments with the highest similarity as the retrieval results, where N is a preset positive integer.
[0010] As a preferred embodiment of the vehicle driving planning method of the present invention, in step S4, the endogenous basic strategy and the historical strategy kernel are fused and arbitrated through a meta-policy network, including the following steps: obtaining the policy feature vector and the historical strategy kernel and its historical performance file contained in the historical strategy memory fragment; analyzing a comprehensive confidence score for the endogenous basic strategy and each historical strategy kernel; the comprehensive confidence score is determined based on the average performance, similarity and performance stability of the endogenous basic strategy in similar historical scenarios; and selecting the strategy with the highest comprehensive confidence score from the endogenous basic strategy and all historical strategy kernels as the final execution strategy.
[0011] As a preferred embodiment of the vehicle driving planning method described in this invention, the method for calculating the comprehensive confidence score is as follows: The average performance of the strategy in similar historical scenarios is normalized to obtain a performance score; the similarity between the scenario descriptor corresponding to the strategy and the current scenario summary descriptor is used as a similarity score; a stability adjustment term is obtained by subtracting the baseline value of performance stability from the performance stability of the strategy in similar historical scenarios; and the comprehensive confidence score of the strategy is obtained by weighted summation of the performance score, similarity score, and stability adjustment term, based on the following formula:
[0012] Where S is the overall confidence score, R is the average performance value of the policy in similar historical scenarios, f(R) is a function for normalizing the average performance value, C is the similarity between the scenario descriptor corresponding to the policy and the current scenario summary descriptor, and V is the performance stability quantification value of the policy in similar historical scenarios. α, β, and γ are the preset performance stability benchmark values, and α, β, and γ are the weighting coefficients of the corresponding terms.
[0013] As a preferred embodiment of the vehicle driving planning method of the present invention, step S6, which updates the strategy memory bank based on the scene summary descriptor, the final execution strategy, and the performance index, includes the following steps: constructing candidate strategy memory fragments based on the current scene summary descriptor, the final execution strategy, and the performance index; analyzing the similarity between the current scene summary descriptor and all historical scene summary descriptors in the strategy memory bank; if there is a historical scene summary descriptor whose similarity to the current scene summary descriptor is greater than a preset first threshold, then performing a knowledge enhancement operation on the candidate strategy memory fragment and the historical strategy memory fragment corresponding to the historical scene summary descriptor; the knowledge enhancement operation includes updating the historical strategy kernel and performance file; if there is no historical scene summary descriptor whose similarity to the current scene summary descriptor is greater than the preset first threshold, then adding the candidate strategy memory fragment as a new historical strategy memory fragment to the strategy memory bank.
[0014] Secondly, the present invention provides a vehicle driving planning system, comprising, The feature extraction module acquires vehicle environmental perception data, processes the environmental perception data, and extracts primary environmental features. The scene cognition module inputs the primary environmental features into the scene cognition network to generate a decision-oriented scene summary descriptor. The scene summary descriptor is used to quantitatively characterize the type and degree of challenge that the current driving scene poses to decision-making. The memory index module, based on the scene summary descriptor, activates the corresponding upper-level decision module in the hierarchical policy network and simultaneously retrieves historical policy memory fragments similar to the scene summary descriptor in the policy memory bank. The fusion module generates an endogenous basic policy based on the activated upper-layer decision module, and fused and arbitrated the endogenous basic policy and the historical policy kernel through a meta-policy network based on the retrieved historical policy memory fragments to generate the final execution policy. The execution module executes the final execution strategy to control vehicle movement and collects performance indicators of strategy execution in real time. The update module updates the policy memory based on the scenario summary descriptor, the final execution strategy, and the performance indicators, thereby completing the closed-loop accumulation of policy knowledge.
[0015] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein the computer program, when executed by the processor, implements any step of the vehicle driving planning method as described in the first aspect of the present invention.
[0016] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the vehicle driving planning method as described in the first aspect of the present invention.
[0017] The beneficial effects of this invention are as follows: It integrates perception, decision-making, and learning into a whole, fundamentally solving the problems of low efficiency, error-proneness, and poor adaptability. It collects raw image and distance data through sensors and immediately performs preliminary processing to identify key elements such as vehicles, pedestrians, and lane lines. It introduces a scene cognition network and, based on the needs of driving tasks, performs in-depth interpretation of raw image and distance data to generate a concise scene summary descriptor, directly quantifying the current scene situation. According to the situation type of the scene summary descriptor, it activates the decision-making method most suitable for handling this type of situation, selecting the solution with the highest confidence from historical and current solutions as the execution instruction, significantly improving the vehicle's decision-making speed, reliability, and adaptability in complex and variable road conditions. Attached Figure Description
[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart of a vehicle driving planning method. Detailed Implementation
[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0023] Reference Figure 1 As one embodiment of the present invention, this embodiment provides a vehicle driving planning method, including the following steps: S1: Acquire vehicle environmental perception data, process the environmental perception data, and extract primary environmental features; S2: Input the primary environmental features into the scene cognition network to generate a decision-oriented scene summary descriptor, which is used to quantitatively characterize the type and degree of challenge that the current driving scene poses to decision-making; S3: Based on the scene summary descriptor, activate the corresponding upper-level decision module in the hierarchical policy network, and simultaneously retrieve historical policy memory fragments similar to the scene summary descriptor in the policy memory bank; S4: Based on the activated upper-level decision-making module, an endogenous basic policy is generated, and based on the retrieved historical policy memory fragments, the endogenous basic policy and the historical policy kernel are fused and arbitrated through the meta-policy network to generate the final execution policy. S5: Execute the final execution strategy to control vehicle movement and collect the performance indicators of strategy execution in real time; S6: Update the strategy memory based on the scenario summary descriptor, the final execution strategy, and the performance indicators to complete the closed-loop accumulation of strategy knowledge.
[0024] As a preferred embodiment of the vehicle driving planning method of the present invention, in step S2, the primary environmental features are standardized and vector-concatenated to construct a global feature vector, which is then input in parallel into multiple attention mechanism calculation modules in the scene cognition network. Each attention mechanism calculation module corresponds to a preset decision challenge dimension. In each attention mechanism calculation module, the correlation weight between the global feature vector and the decision challenge dimension corresponding to the attention mechanism calculation module is analyzed. The correlation weight represents the degree of influence of elements related to that dimension in the current environment on driving decisions. The correlation weight output by each attention mechanism calculation module is normalized to generate a quantitative scoring sub-vector corresponding to the decision challenge dimension. All quantitative scoring sub-vectors of the decision challenge dimensions are combined to generate the final scene summary descriptor.
[0025] Feature standardization employs Z-score normalization to ensure each feature dimension has zero mean and unit variance. Vector concatenation connects features from different sensors in a predetermined order to form a global feature vector. The scene cognition network includes six parallel attention computation modules, each corresponding to a specific decision challenge dimension. Attention computation uses a scaled dot product attention mechanism. The query vector comes from global features, while the key and value vectors are generated from a learnable parameter matrix. The attention weight calculation process includes three sub-steps: dot product, scaling, and softmax normalization. Normalization uses the Min-Max method to linearly transform the original attention score to the 0-1 range. The final feature combination uses vector concatenation to merge the scores of each dimension into a complete scene summary descriptor.
[0026] As a preferred embodiment of the vehicle driving planning method described in this invention, the relevance weight analysis process in the relevance weight of the global feature vector and the decision challenge dimension corresponding to the attention mechanism calculation module includes: generating a query vector from the global feature vector through linear transformation; generating a key vector from the feature representation of the decision challenge dimension through linear transformation; performing a dot product operation between the query vector and the key vector and then processing it through a nonlinear activation function to obtain the original relevance score of the decision challenge dimension; and normalizing the original relevance score to generate the relevance weight.
[0027] Linear transformation operations are implemented using fully connected layers, with the query and key vectors set to 64 dimensions. The dot product operation calculates the inner product of the query and key vectors to obtain the original relevance score. Scaling divides the score by the square root of the vector dimension to prevent the vanishing gradient problem. The ReLU function is used as the non-linear activation function to enhance the model's non-linear expressiveness. Layer Normalization is used to standardize the output of each attention head. The final relevance weights are calculated using the Softmax function to ensure that the sum of all weights is 1.
[0028] As a preferred embodiment of the vehicle driving planning method of the present invention, wherein: in step S3, retrieving historical strategy memory fragments in the strategy memory bank based on scene summary descriptors includes the following sub-steps: obtaining multiple historical strategy memory fragments stored in the strategy memory bank, each historical strategy memory fragment containing a historical scene summary descriptor; analyzing the similarity between the current scene summary descriptor and each historical scene summary descriptor using the cosine similarity method; sorting the historical strategy memory fragments in descending order of similarity; and selecting the top N historical strategy memory fragments with the highest similarity as the retrieval results, where N is a preset positive integer.
[0029] The strategy memory uses a graph database to store historical records. Each node contains a scene descriptor, strategy parameters, and performance metrics. Similarity calculation employs an improved cosine similarity algorithm, introducing weight coefficients to adjust the importance of different dimensions. The sorting algorithm uses quicksort, with a time complexity of O(nlogn). The number of search results, N, is dynamically adjusted based on the size of the memory, and the optimal value of N is determined using the square root rule. The retrieval process includes four sub-steps: feature extraction, similarity calculation, result sorting, and result truncation.
[0030] As a preferred embodiment of the vehicle driving planning method of the present invention, in step S4, the endogenous basic strategy and the historical strategy kernel are fused and arbitrated through a meta-policy network, including the following steps: obtaining the policy feature vector and the historical strategy kernel and its historical performance file contained in the historical strategy memory fragment; analyzing a comprehensive confidence score for the endogenous basic strategy and each historical strategy kernel; the comprehensive confidence score is determined based on the average performance, similarity and performance stability of the endogenous basic strategy in similar historical scenarios; and selecting the strategy with the highest comprehensive confidence score from the endogenous basic strategy and all historical strategy kernels as the final execution strategy.
[0031] The strategy feature vector extraction adopts an autoencoder architecture to compress the strategy parameters into a low-dimensional representation. The historical performance profile includes statistical indicators such as success rate and average reward value. The comprehensive confidence score calculation introduces a time decay factor, giving higher weight to recent performance. The strategy selection adopts an ε-greedy strategy, which selects the optimal strategy with a higher probability and explores the suboptimal strategy with a lower probability. The arbitration process includes four stages: feature extraction, score calculation, strategy selection, and parameter adjustment.
[0032] As a preferred embodiment of the vehicle driving planning method described in this invention, the method for calculating the comprehensive confidence score is as follows: The average performance of the strategy in similar historical scenarios is normalized to obtain a performance score; the similarity between the scenario descriptor corresponding to the strategy and the current scenario summary descriptor is used as a similarity score; a stability adjustment term is obtained by subtracting the baseline value of performance stability from the performance stability of the strategy in similar historical scenarios; and the comprehensive confidence score of the strategy is obtained by weighted summation of the performance score, similarity score, and stability adjustment term, based on the following formula:
[0033] Where S is the overall confidence score, R is the average performance value of the policy in similar historical scenarios, f(R) is a function for normalizing the average performance value, C is the similarity between the scenario descriptor corresponding to the policy and the current scenario summary descriptor, and V is the performance stability quantification value of the policy in similar historical scenarios. α, β, and γ are the preset performance stability benchmark values, and α, β, and γ are the weighting coefficients of the corresponding terms.
[0034] The average performance value under similar historical scenarios reflects the past performance of the strategy; the scenario matching term measures the suitability of the strategy to the current scenario; the stability term assesses the volatility of the strategy's performance; and the weight coefficients adjust the importance of each term. The historical performance term is normalized using the sigmoid function, mapping the original reward value to the 0-1 range; the scenario matching term directly uses the cosine similarity calculation result; the stability term calculates the variance of the strategy's performance in historical scenarios; and the weight coefficients determine the optimal value through grid search, balancing the strategy's performance, suitability, and stability.
[0035] As a preferred embodiment of the vehicle driving planning method of the present invention, step S6, which updates the strategy memory bank based on the scene summary descriptor, the final execution strategy, and the performance index, includes the following steps: constructing candidate strategy memory fragments based on the current scene summary descriptor, the final execution strategy, and the performance index; analyzing the similarity between the current scene summary descriptor and all historical scene summary descriptors in the strategy memory bank; if there is a historical scene summary descriptor whose similarity to the current scene summary descriptor is greater than a preset first threshold, then performing a knowledge enhancement operation on the candidate strategy memory fragment and the historical strategy memory fragment corresponding to the historical scene summary descriptor; the knowledge enhancement operation includes updating the historical strategy kernel and performance file; if there is no historical scene summary descriptor whose similarity to the current scene summary descriptor is greater than the preset first threshold, then adding the candidate strategy memory fragment as a new historical strategy memory fragment to the strategy memory bank.
[0036] Candidate memory fragments are formatted in a structured data format, including timestamps, scene features, strategy parameters, and performance metrics. A threshold of 0.8 ensures that only highly similar records trigger knowledge fusion. Knowledge reinforcement employs a moving average algorithm, with a weight ratio of 1:9 between new and old records. New records undergo integrity verification, and records missing key fields are discarded. The update process comprises four stages: fragment construction, similarity calculation, knowledge fusion, and record addition. A regular cleanup mechanism maintains the memory bank, removing expired or low-quality records.
[0037] This embodiment also provides a vehicle driving planning system, including: The feature extraction module acquires vehicle environmental perception data, processes the environmental perception data, and extracts primary environmental features. The scene cognition module inputs the primary environmental features into the scene cognition network to generate a decision-oriented scene summary descriptor. The scene summary descriptor is used to quantitatively characterize the type and degree of challenge that the current driving scene poses to decision-making. The memory index module, based on the scene summary descriptor, activates the corresponding upper-level decision module in the hierarchical policy network and simultaneously retrieves historical policy memory fragments similar to the scene summary descriptor in the policy memory bank. The fusion module generates an endogenous basic policy based on the activated upper-layer decision module, and fused and arbitrated the endogenous basic policy and the historical policy kernel through a meta-policy network based on the retrieved historical policy memory fragments to generate the final execution policy. The execution module executes the final execution strategy to control vehicle movement and collects performance indicators of strategy execution in real time. The update module updates the policy memory based on the scenario summary descriptor, the final execution strategy, and the performance indicators, thereby completing the closed-loop accumulation of policy knowledge.
[0038] This embodiment also provides a computer device applicable to vehicle driving planning methods, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the vehicle driving planning method proposed in the above embodiment.
[0039] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0040] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the vehicle driving planning method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0041] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A vehicle driving planning method, characterized in that: include, S1: Acquire vehicle environmental perception data, process the environmental perception data, and extract primary environmental features; S2: Input the primary environmental features into the scene cognition network to generate a decision-oriented scene summary descriptor, which is used to quantitatively characterize the type and degree of challenge that the current driving scene poses to decision-making; S3: Based on the scene summary descriptor, activate the corresponding upper-level decision module in the hierarchical policy network, and simultaneously retrieve historical policy memory fragments similar to the scene summary descriptor in the policy memory bank; S4: Based on the activated upper-level decision-making module, an endogenous basic policy is generated, and based on the retrieved historical policy memory fragments, the endogenous basic policy and the historical policy kernel are fused and arbitrated through the meta-policy network to generate the final execution policy. S5: Execute the final execution strategy to control vehicle movement and collect the performance indicators of strategy execution in real time; S6: Update the strategy memory based on the scenario summary descriptor, the final execution strategy, and the performance indicators to complete the closed-loop accumulation of strategy knowledge.
2. The vehicle driving planning method as described in claim 1, characterized in that: In step S2, the primary environmental features are standardized and concatenated to construct a global feature vector, which is then input into multiple attention mechanism calculation modules in the scene cognition network in parallel. Each attention mechanism calculation module corresponds to a preset decision challenge dimension. In each attention mechanism calculation module, the correlation weight between the global feature vector and the decision challenge dimension corresponding to the attention mechanism calculation module is analyzed. The correlation weight represents the degree of influence of elements related to that dimension in the current environment on driving decisions. The correlation weight output by each attention mechanism calculation module is normalized to generate a quantitative scoring sub-vector for the corresponding decision challenge dimension. All quantitative scoring sub-vectors for decision challenge dimensions are combined to generate the final scene summary descriptor.
3. The vehicle driving planning method as described in claim 2, characterized in that: In the relevance weights of the decision challenge dimension corresponding to the global feature vector and attention mechanism calculation module, the relevance weight analysis process includes: generating a query vector from the global feature vector through linear transformation; generating a key vector from the feature representation of the decision challenge dimension through linear transformation; performing a dot product operation between the query vector and the key vector and then processing it through a nonlinear activation function to obtain the original relevance score of the decision challenge dimension; and normalizing the original relevance score to generate the relevance weights.
4. The vehicle driving planning method as described in claim 3, characterized in that: In step S3, retrieving historical strategy memory fragments from the strategy memory bank based on scene summary descriptors includes the following steps: obtaining multiple historical strategy memory fragments stored in the strategy memory bank, each historical strategy memory fragment containing a historical scene summary descriptor; analyzing the similarity between the current scene summary descriptor and each historical scene summary descriptor using the cosine similarity method; sorting the historical strategy memory fragments in descending order of similarity; and selecting the top N historical strategy memory fragments with the highest similarity as the retrieval results, where N is a preset positive integer.
5. The vehicle driving planning method as described in claim 4, characterized in that: In step S4, the endogenous basic strategy and historical strategy kernel are fused and arbitrated through a meta-policy network, including the following steps: obtaining the policy feature vector and the historical strategy kernel and its historical performance file contained in the historical strategy memory fragment; analyzing a comprehensive confidence score for the endogenous basic strategy and each historical strategy kernel; the comprehensive confidence score is determined based on the average performance, similarity and performance stability of the endogenous basic strategy in similar historical scenarios; and selecting the strategy with the highest comprehensive confidence score from the endogenous basic strategy and all historical strategy kernels as the final execution strategy.
6. The vehicle driving planning method as described in claim 5, characterized in that: The method for calculating the comprehensive confidence score is as follows: The average performance of the policy in similar historical scenarios is normalized to obtain the performance score. The similarity between the scenario descriptor corresponding to the policy and the current scenario summary descriptor is used as the similarity score. The baseline value of performance stability is subtracted from the performance stability of the policy in similar historical scenarios to obtain the stability adjustment term. The comprehensive confidence score of the policy is obtained by weighted summation of the performance score, similarity score, and stability adjustment term. The formula used is: Where S is the overall confidence score, R is the average performance value of the policy in similar historical scenarios, f(R) is a function for normalizing the average performance value, C is the similarity between the scenario descriptor corresponding to the policy and the current scenario summary descriptor, and V is the performance stability quantification value of the policy in similar historical scenarios. α, β, and γ are the preset performance stability benchmark values, and α, β, and γ are the weighting coefficients of the corresponding terms.
7. The vehicle driving planning method as described in claim 6, characterized in that: The step S6, which updates the policy memory based on the scene summary descriptor, the final execution strategy, and the performance index, includes the following steps: constructing candidate policy memory fragments based on the current scene summary descriptor, the final execution strategy, and the performance index; analyzing the similarity between the current scene summary descriptor and all historical scene summary descriptors in the policy memory; if there is a historical scene summary descriptor whose similarity to the current scene summary descriptor is greater than a preset first threshold, then performing a knowledge enhancement operation on the candidate policy memory fragment and the historical policy memory fragment corresponding to that historical scene summary descriptor. The knowledge enhancement operation includes updating the historical policy kernel and performance profile; if there is no historical scene summary descriptor whose similarity to the current scene summary descriptor is greater than the preset first threshold, then adding the candidate policy memory fragment as a new historical policy memory fragment to the policy memory.
8. A vehicle driving planning system, based on the vehicle driving planning method according to any one of claims 1 to 7, characterized in that: include, The feature extraction module acquires vehicle environmental perception data, processes the environmental perception data, and extracts primary environmental features. The scene cognition module inputs the primary environmental features into the scene cognition network to generate a decision-oriented scene summary descriptor, which is used to quantitatively characterize the type and degree of challenge that the current driving scene poses to decision-making. The memory index module, based on the scene summary descriptor, activates the corresponding upper-level decision module in the hierarchical policy network and simultaneously retrieves historical policy memory fragments similar to the scene summary descriptor in the policy memory bank. The fusion module generates an endogenous basic policy based on the activated upper-level decision module, and fused and arbitrated the endogenous basic policy and the historical policy kernel through a meta-policy network based on the retrieved historical policy memory fragments to generate the final execution policy. The execution module executes the final execution strategy to control vehicle movement and collects performance indicators of strategy execution in real time. The update module updates the policy memory based on the scenario summary descriptor, the final execution strategy, and the performance indicators, thereby completing the closed-loop accumulation of policy knowledge.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the vehicle driving planning method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the vehicle driving planning method according to any one of claims 1 to 7.