Self-adaptive sensing strategy switching method and system for long-distance automatic driving

Abstract

The invention relates to a self-adaptive perception strategy switching method and system for long-distance automatic driving, and the method comprises the steps: carrying out the analysis of a public data set through employing a multi-target joint optimization clustering method, generating a standard scene prototype, and constructing an initial scene-optimal strategy mapping database; acquiring multi-dimensional scene features of a navigation route, performing global optimal scene division and recognition on a target road section, and generating a perception strategy map between continuous cells; querying a scene-optimal strategy mapping database continuously optimized by reinforcement learning, and selecting an optimal perception strategy for each cell based on a Q value maximization principle; when the vehicle enters the new cell, the self-adaptive switching control module automatically activates the preloaded optimal sensing strategy; and evaluating a strategy execution effect, generating an instant reward, transmitting an experience tuple containing a state, an action and the reward to the cloud, and continuously optimizing the Q-value database. The method and the system can improve the perception performance and the resource utilization efficiency of the automatic driving system.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to an adaptive perception strategy switching method and system for long-distance autonomous driving. Background Technology

[0002] With the rapid development of intelligent transportation and autonomous driving technologies, vehicle environmental perception systems have become a core module for achieving safe driving and intelligent decision-making. Mainstream autonomous vehicles are typically equipped with multiple sensors (such as cameras, LiDAR, and millimeter-wave radar) and employ technologies such as multi-sensor fusion and deep learning perception to achieve high-precision detection and identification of targets such as roads, obstacles, and pedestrians. These perception systems provide essential environmental information support for autonomous vehicles and are fundamental to ensuring the safe and reliable operation of autonomous driving.

[0003] Current autonomous driving perception systems generally employ a fixed perception strategy, meaning they continuously run the same type of perception model regardless of changes in the actual traffic scenario. This deficiency is particularly pronounced in long-distance autonomous driving tasks, where a single long-distance journey often traverses multiple complex and vastly different environments. This makes it difficult for the system to meet the perception needs of various scenarios, including different road types, traffic densities, weather conditions, and special situations. For example, while multi-sensor fusion strategies offer strong robustness, the surge in data processing volume and system computational latency in high-density traffic environments significantly impacts real-time decision-making. Furthermore, in scenarios with numerous small targets, the sparsity of LiDAR point clouds becomes a prominent issue, leading to decreased detection accuracy. While pure vision models perform well in simple environments such as structured environments and daytime conditions, their perception performance deteriorates severely in complex scenarios such as high occlusion, extreme weather, and high-density traffic flow, making them prone to blind spots, false detections, and missed detections.

[0004] Some autonomous driving solutions have introduced perception strategy switching mechanisms, which dynamically adjust perception models or sensor fusion methods based on certain environmental characteristics. For example, Baidu Apollo automatically prioritizes LiDAR or adjusts fusion weights in extreme weather conditions such as rain, snow, and fog, and switches to specific perception models at night or in tunnel scenarios. These solutions typically switch strategies through static rules or limited scene classification, while some researchers have also proposed methods for adjusting sensor configurations based on environmental variables.

[0005] Despite these shortcomings, existing perception strategy switching methods still have significant limitations. First, current switching mechanisms are mostly passive, responding to immediate environmental characteristics without considering the global and forward-looking aspects of long-distance route planning. Second, most existing methods only support simple switching at the sensor or weight level, failing to achieve intelligent adaptive adjustment of perception models or strategies and struggling to cope with the changing demands of complex traffic environments. Furthermore, existing systems generally lack a scientific "scene-optimal strategy" mapping mechanism based on large-scale data; the strategy selection process lacks theoretical basis, exhibiting blindness and lag. While some engineering applications support manual switching of specific models in extreme scenarios, the switching conditions are simple, lacking systematic stability and efficiency guarantees. Therefore, there is an urgent need for an adaptive perception strategy switching method capable of high-precision scene classification and scientific decision-making to improve the perception performance and resource utilization efficiency of autonomous driving systems in variable traffic environments. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive perception strategy switching method and system for long-distance autonomous driving, thereby improving the perception performance and resource utilization efficiency of the autonomous driving system.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: an adaptive perception strategy switching method for long-distance autonomous driving, comprising the following steps:

[0008] S1: The public dataset is analyzed using a multi-objective joint optimization clustering method to generate standard scenario prototypes, and then an initial scenario-optimal strategy mapping database is constructed.

[0009] S2: Obtain the environmental, traffic, and dynamic target multi-dimensional scene features of the complete planned navigation route, perform global optimal scene division and identification of the target road segment, and generate a perception strategy map containing a series of continuous cells with scene labels.

[0010] S3: Query the scene-optimal policy mapping database continuously optimized by reinforcement learning, select the optimal perception policy for each interval based on the Q-value maximization principle, and trigger the preloading module;

[0011] S4: By monitoring the vehicle's position, the adaptive switching control module automatically activates the pre-loaded optimal perception strategy when the vehicle enters a new small area.

[0012] S5: Evaluate the effectiveness of strategy execution and generate immediate rewards. Upload experience tuples containing status, actions, and rewards to the cloud and continuously optimize the Q-value database used through a temporal difference learning algorithm.

[0013] Furthermore, in step S1, a multi-objective joint optimization clustering method is first used on the cloud server to generate a set of statistically optimal standard scene categories from scene fragments extracted from the public dataset, and an initial knowledge base is constructed based on these categories; the core is to find a set C of optimal clustering scheme clusters. * With feature subset F * Its expression is:

[0014] (C * ,F * ) = argmin C,F J(C,F),

[0015] J(C,F)=αD intra (C)+βO inter (C)+γR feat (F)

[0016] Among them, D intra (C) represents the intra-cluster dispersion, O inter (C) represents the inter-cluster overlap, R feat (F) represents feature effectiveness, where α, β, and γ are adjustable hyperparameters used to balance the importance of different optimization objectives;

[0017] A genetic algorithm is used to search for an approximate optimal solution to the above optimization problem. First, a clustering scheme (C,F) is encoded into a chromosome G. The chromosome contains an integer vector of the cluster label to which each data point belongs and a binary mask of the selected feature subset. Then, a clustering scheme containing N is randomly generated. pop Initial populations of different chromosomes For each chromosome G in the population j Calculate its corresponding fitness score, the formula is as follows:

[0018]

[0019] Among them, J(C j ,F j ) represents the joint optimization function value corresponding to the decoded chromosome, and ε is a very small normal number;

[0020] The population is optimized by repeatedly performing a series of evolutionary operations; in each generation, a roulette wheel selection strategy is first used, based on the fitness score (G). j The chromosome is probabilistically selected from the current population for reproduction; the probability p of each chromosome being selected is... j It is proportional to its fitness, and its expression is:

[0021]

[0022] Then, with a preset crossover probability p c The selected parents are paired and their gene segments are exchanged to generate new offspring chromosomes; finally, with a mutation probability p... m The newly generated offspring chromosomes are randomly perturbed to maintain population diversity; the iterative evolution process is repeated until the preset maximum number of iterations T is reached. max The algorithm terminates when the fitness of the best individual in the population no longer significantly improves over multiple generations; at this point, the chromosome with the highest fitness in the current population is considered the final approximate optimal solution (Ci). * ,F * );

[0023] Each final cluster is defined as a standard scene prototype, forming a standard scene prototype library. Based on the results of systematic comparative testing experiments on various perception strategies under these standard scene categories, an initial Q value Q(s,a) is assigned to each scene prototype (state s) and each available perception strategy (action a). All initial Q values ​​together constitute the initial scene-optimal strategy mapping database.

[0024] Furthermore, in step S2, firstly, for the predetermined long-distance target driving route, multi-source data is fused to construct an instantaneous, multi-dimensional scene feature stream; then, the continuous scene feature stream is scanned, and any feature that undergoes a significant change is marked as a potential change point, generating a set P = {p1, p2, ..., p...} of N ordered candidate segmentation points. N}; then for any two candidate points p j and p i For each defined potential subinterval seg(j,i), a unique interval matching cost is calculated, and the best standard scene cluster it matches is recorded, 1≤j≤i≤N; a set of M standard scene clusters C={c1,c2,...,c M In the library, the standard scene cluster that is closest to μ(seg(j,i)) is found through nearest neighbor matching. Calculate the distance from all points within the potential interval to the centroid of this optimal standard cluster. The average squared distance is used as its matching cost C. match (seg(j,i));

[0025] Then, a dynamic programming algorithm is used to find the optimal K-1 final split points among N candidate split points; a two-dimensional table Cost[i][k] is constructed to store the minimum total cost of dividing the road segment covered by the first i candidate points into k cells;

[0026] After the dynamic programming algorithm has finished executing and filled the entire Cost table, the optimal segmentation scheme for the specific route can be obtained by backtracking Cost[N][K]. This series of small intervals and labels constitutes the perception strategy map for this long-distance task.

[0027] Furthermore, in step S3, based on the pre-built scenario-optimal strategy mapping database and a data-driven intelligent decision-making model, the optimal perception strategy is matched and selected for each cell, and the corresponding optimal perception strategy is pre-loaded into the vehicle's local system. Specifically, this includes:

[0028] When selecting a policy between cells, the scenario-optimal policy mapping module queries the scenario-optimal policy mapping database, which is continuously optimized by the reinforcement learning policy optimization module. For a standard scenario type (state s) that has been classified in step S2, the intelligent decision model selects from the set A = {a1, a2, ..., a...} of all available perception policies corresponding to that scenario. n In the given information, select strategy a that has the highest expected Q value. * ;

[0029] Determining the optimal strategy a * Then, the strategy preloading module is triggered to preload the perception model and related configuration files required for the optimal strategy into the active memory of the computing unit. When the vehicle enters the small area, the adaptive switching control module activates the strategy.

[0030] Furthermore, in step S4, during vehicle operation, the vehicle continuously locates its current position; upon entering a new inter-cell area, it automatically switches and activates a perception model and parameter configuration matching the scenario type, specifically including:

[0031] During actual vehicle operation, the onboard high-precision positioning module continuously outputs the vehicle's precise geographical coordinates and heading information by fusing the matching results of the Global Navigation Satellite System and high-precision maps. The adaptive switching control module continuously compares this real-time location with the perception strategy map. When the system detects that the vehicle's current location has crossed the boundary of the next cell, a switching event is triggered. At this time, the adaptive switching control module immediately sends a control command containing a new strategy identifier to the adaptive perception execution module. Upon receiving the command, the perception system execution unit immediately activates the perception model and corresponding parameter configuration that are pre-loaded into memory by the strategy pre-loading module and match the scene type between the new cells. Simultaneously, the system unloads or suspends the perception model of the previous cell to release computing resources.

[0032] Furthermore, in step S5, the scene features, strategy selection, and switching result data are recorded in local and cloud databases to continuously optimize the scene classification and strategy mapping model parameters, specifically including:

[0033] After the strategy is activated and executed for a period of time, the vehicle-mounted performance evaluation and reward generation module is responsible for quantitatively evaluating the actual performance of the current perception strategy and generating an instant reward value r.

[0034] Then, the data recording and uploading module packages the complete information constituting this decision-making and execution experience into an experience tuple (s,a,r,s') and uploads it to the cloud-based strategy generation and optimization platform, where s is the scenario category before the strategy is executed, a is the actual perception strategy executed, r is the just calculated immediate reward, and s' is the new scenario category where the vehicle is currently located; the reinforcement learning strategy optimization module uses this data to continuously optimize the Q-value stored in the scenario-optimal strategy mapping database.

[0035] The present invention also provides an adaptive perception strategy switching system for long-distance autonomous driving, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the above-mentioned method.

[0036] Compared with existing technologies, this invention has the following advantages: Addressing the problem that existing autonomous driving perception systems struggle to flexibly switch optimal perception strategies based on changing traffic scenarios when performing long-distance tasks, resulting in insufficient perception accuracy and system efficiency under different road types, traffic densities, weather conditions, and special scenarios, this invention proposes an adaptive perception strategy switching method for long-distance autonomous driving. This method utilizes a multi-objective joint optimization clustering approach to offline construct a high-quality "standard scenario prototype library" from a large-scale dataset. Based on prior maps, traffic flow, weather, and other multi-source data, long-distance road segments are divided into smaller intervals using a dynamic programming-based optimal sequence segmentation algorithm and categorized into standard scenario types. For each segmented interval, based on a Q-value database continuously optimized by reinforcement learning, the perception strategy most suitable for that scenario type is intelligently preloaded and automatically activated upon reaching the corresponding interval. This mechanism enables the vehicle to automatically select and switch to a more robust multi-sensor fusion model when facing typical complex scenarios such as high occlusion, high-density targets, and severe weather, thereby improving the accuracy and stability of target detection. Simultaneously, in simple, structured scenarios with high visibility, it can automatically switch to a low-power, high-efficiency pure vision model. This dynamically balances perception performance and computational resource utilization while ensuring safety, effectively avoiding performance bottlenecks and resource waste associated with single perception schemes under specific conditions. This invention not only achieves scene-adaptive intelligent strategy switching but also reduces system latency and energy consumption caused by frequent model loading, providing a solid theoretical foundation and engineering support for the safe, reliable, and efficient operation of autonomous vehicles in variable traffic environments. Attached Figure Description

[0037] Figure 1 The system model diagram corresponding to the adaptive perception strategy switching method for long-distance autonomous driving provided in the embodiments of the present invention;

[0038] Figure 2 A flowchart of an adaptive perception strategy switching method for long-distance autonomous driving provided in an embodiment of the present invention. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] To address the safety and efficiency issues of existing autonomous driving perception systems, which rely on fixed strategies and struggle to adapt to changing driving environments, this invention proposes the first autonomous driving adaptive perception strategy switching method that integrates prior knowledge and adaptive learning. This method is based on multi-objective joint optimization clustering technology, initial mapping relationships constructed from large-scale experimental data, and reinforcement learning adaptive optimization technology. Figure 1 As shown, the implementation of this method mainly comprises two categories of entities: a policy generation and optimization platform and a perception policy execution and data acquisition system. First, a multi-objective joint optimization clustering method is used to perform deep analysis on a large-scale public dataset to generate a high-quality "standard scene prototype library." Combined with comparative testing experiments, a "scene-optimal policy" mapping database containing initial Q-values ​​is constructed. After the vehicle plans a specific route, the real-time multi-dimensional scene features of that route are acquired. Using a dynamic programming-based optimal sequence segmentation algorithm and the constructed prototype library, an optimal policy map containing optimal road segmentation and scene recognition results is "tailor-made" for that route. During vehicle operation, the perception policy execution and data acquisition system obtains this policy map from the cloud platform, and the policy preloading module prepares the perception model in advance. The high-precision positioning module provides the vehicle's real-time location, and the adaptive switching control module triggers the corresponding policy activation based on this location information. Simultaneously, the performance evaluation and reward generation module evaluates the policy execution effect and generates reward values. Finally, the data recording and uploading module uploads the experience tuples to the cloud platform, where the reinforcement learning policy optimization module continuously updates the Q-values ​​in the "scenario-optimal policy" mapping database, thus forming a complete adaptive learning and evolutionary closed loop.

[0043] like Figure 2 As shown, the adaptive perception strategy switching method for long-distance autonomous driving provided in this embodiment includes the following steps:

[0044] S1: Use a multi-objective joint optimization clustering method to analyze the public dataset, generate standard scenario prototypes, and then build an initial "scenario-optimal strategy" mapping database;

[0045] S2: Obtain multi-dimensional scene features such as environment, traffic, and dynamic targets for the complete planned navigation route, perform global optimal scene division and recognition on the target road segment, and generate a "perception strategy map" containing a series of continuous cells with scene labels.

[0046] S3: Query the "scene-optimal policy" mapping database continuously optimized by reinforcement learning, select the optimal perception policy for each interval based on the Q-value maximization principle, and trigger the preloading module;

[0047] S4: By monitoring the vehicle's position through high-precision positioning, the adaptive switching control module automatically activates the pre-loaded optimal perception strategy when the vehicle enters a new small area.

[0048] S5: Evaluate the effectiveness of strategy execution and generate immediate rewards. Upload experience tuples containing status, actions, and rewards to the cloud and continuously optimize the Q-value database used through a temporal difference learning algorithm.

[0049] In step S1, a multi-objective joint optimization clustering method is first used on a cloud server to generate a set of statistically optimal standard scene categories from massive scene fragments extracted from large-scale public datasets nuScenes and KITTI. An initial knowledge base is then constructed based on these categories. The core objective is to find a set C of optimal clustering scheme clusters. * With feature subset F * Its expression is:

[0050] (C * ,F * ) = argmin C,F J(C,F),

[0051] J(C,F)=αD intra (C)+βO inter (C)+γR feat (F)

[0052] Among them, D intra (C) represents the intra-cluster dispersion, O inter (C) represents the inter-cluster overlap, R feat (F) represents feature effectiveness, where α, β, and γ are adjustable hyperparameters used to balance the importance of different optimization objectives. The specific definitions of each component of this function are as follows.

[0053] Intra-cluster dispersion D intra (C) is used to quantify the dispersion of scene features within all clusters, aiming to ensure high homogeneity within each scene category. A smaller value indicates a more compact cluster, and its expression is:

[0054]

[0055] Where C = {c1, c2, ..., c} k Let x be a scheme that divides all scene fragments into K clusters. i Belongs to cluster c k A scene feature vector, while μ k Then w is the centroid of the cluster. i To compare with sample x i The relevant sample weights, W, are used to characterize the representativeness or confidence level of the sample. k This is a feature weighting matrix used to define the importance of different feature dimensions when calculating intra-cluster dispersion.

[0056] Inter-cluster overlap O inter (C) is used to quantify the boundary ambiguity between different clusters, with the goal of maximizing the distinguishability between different scene categories. Its expression is:

[0057]

[0058] Where, p k (x) and p j (x) represent clusters c k and c j The probability density function is a multivariate Gaussian distribution. μ and Σ are the mean vector and covariance matrix of the cluster, respectively.

[0059] Feature effectiveness R feat (F) is used to quantify the information redundancy of the currently selected feature subset F in expressing scene differences. Its goal is to enable the algorithm to select a feature combination with high information content but low internal correlation. Normalized average mutual information is used to measure this redundancy; the higher the value, the more redundant information exists between features. Its expression is:

[0060]

[0061] Where F is the feature subset used in the current scheme; |F| is the size of the subset; double summation ΣΣ represents calculating the sum of mutual information between all distinct feature pairs in the subset, I N (f i ;f j ) is a feature f i and feature f j Normalized mutual information between them.

[0062] Due to the complexity of this optimization problem, the system employs a genetic algorithm to search for its approximate optimal solution. The algorithm first encodes a complete clustering scheme (C, F) into a "chromosome" G. The chromosome contains an integer vector of the cluster label to which each data point belongs and a binary mask of the selected feature subset. Then, the system randomly generates a cluster containing N... pop Initial populations of different chromosomes For each chromosome G in the population j The algorithm will calculate the corresponding fitness score, and the calculation formula is as follows:

[0063]

[0064] Among them, J(C j ,F j ) represents the joint optimization function value corresponding to the decoded chromosome, and ε is a very small normal number.

[0065] The algorithm optimizes the population by iteratively executing a series of evolutionary operations. In each generation, a roulette wheel selection strategy is first employed, based on the fitness score (G). j The chromosome is probabilistically selected from the current population to reproduce, with each chromosome having a probability p of being selected. j It is proportional to its fitness, and its expression is:

[0066]

[0067] Then, with a preset crossover probability p c The selected parent chromosomes are paired, and their gene segments are exchanged through single-point or multi-point crossover to generate new "offspring" chromosomes. Finally, with a low mutation probability p... m The newly generated offspring chromosomes are randomly perturbed to maintain population diversity. This iterative evolutionary process is repeated until the preset maximum number of iterations T is reached. max The algorithm terminates when the fitness of the best individual in the population no longer significantly improves over multiple generations. At this point, the chromosome with the highest fitness in the current population is considered the final approximate optimal solution (Ci). * ,F * ).

[0068] The system defines each final cluster as a standard scene prototype, and these prototypes ultimately form a "standard scene prototype library". At the same time, the system combines the results of systematic comparative testing experiments on various perception strategies under these standard scene categories, and assigns an initial Q value Q(s,a) to each scene prototype (state s) and each available perception strategy (action a). All initial Q values ​​together constitute the initial "scene-optimal strategy" mapping database.

[0069] In step S2, the system first fuses multi-source data for a predetermined long-distance target driving route to construct an instantaneous, multi-dimensional scene feature stream. Then, the system scans this continuous scene feature stream, and any significant change in a feature is marked as a potential change point, generating a set P = {p1, p2, ..., p...} of N ordered candidate segmentation points. N}; then for any two candidate points p j and p i For each defined potential subinterval seg(j,i), a unique interval matching cost is calculated, and the best standard scene cluster it matches is recorded, 1≤j≤i≤N; a set of M standard scene clusters C={c1,c2,...,c M In the library, the standard scene cluster that is closest to μ(seg(j,i)) is found through nearest neighbor matching. Calculate the distance from all points within the potential interval to the centroid of this optimal standard cluster. The average squared distance is used as its matching cost C. match (seg(j,i)). Its expression is:

[0070]

[0071] Where, μ m For standard scene cluster c m The centroid of point p. F(p) represents the normalized feature vector of point p. All calculated matching costs are stored in a two-dimensional matrix MatchCostMatrix[j][i]. At the same time, the best standard cluster index m* matched by each potential interval seg(j,i) is also recorded in another matrix BestClusterIndex[j][i].

[0072] Then, a dynamic programming algorithm is used to find the optimal K-1 final split points among the N candidate split points. The algorithm constructs a two-dimensional table Cost[i][k] to store the minimum total cost of dividing the road segment covered by the first i candidate points into k cells. This table is filled with the following recursive relation:

[0073]

[0074] Among them, C seg (j,i) represents the candidate point p. j to p i The total cost incurred by this segment of the path as a single small interval. This cost consists of its own matching cost and the switching cost with the previous interval, and its expression is:

[0075] C seg(j,i)=MatchCostMatrix[j+1][i]+λ·C switch (s k-1 ,seg(j+1,i))

[0076] Among them, the scene switching cost C switch The "cost" used to quantify switching from the scenario type of the previous optimal interval to the scenario type of the current new evaluation interval is expressed as follows:

[0077]

[0078] Here, ε is a very small positive constant to avoid the denominator being zero. and To obtain the previous optimal interval s by querying the BestClusterIndex matrix. k-1 The best standard cluster index that matches the current new interval seg(j+1,i).

[0079] After the dynamic programming algorithm has finished executing and the entire Cost table has been filled, the optimal segmentation scheme for that specific route can be obtained by backtracking on Cost[N][K]. This scheme not only clarifies the precise locations of the final selected K-1 segmentation points, but also provides the optimal interval s for each generated segment. k Matched standard scene clusters The tags. This series of small intervals and tags constitutes the "perception strategy map" for this long-distance mission.

[0080] In step S3, based on the pre-built "scene-perception strategy" mapping table and a data-driven intelligent decision-making model, the optimal perception strategy is matched and selected for each cell, and the corresponding optimal perception strategy is pre-loaded into the vehicle's local system. Specifically, this includes:

[0081] When selecting a strategy between cells, the "scenario-optimal strategy" mapping module queries a dynamic "scenario-optimal strategy" mapping database that is continuously optimized by the reinforcement learning strategy optimization module. The initial mapping relationships in this database are constructed from empirical conclusions derived from systematic comparative testing experiments of various perception strategies in diverse scenarios. For a specific "standard scenario type" (state s) that has been classified in step S2, the intelligent decision model will select from the set A = {a1, a2, ..., a...} of all available perception strategies corresponding to that scenario. n In the given information, select strategy a that has the highest expected Q value. * The expression for this decision-making process is:

[0082] a * =argmax a∈AQ(s,a)

[0083] Among them, a * Let be the selected optimal strategy; A represents the set of all available perception strategies in the system, mainly including pure visual perception strategies, multi-sensor fusion strategies, and Transformer-based fusion strategies; Q(s,a) is a quantitative estimate of the expected long-term reward of performing action a in state s, retrieved from this dynamic database. By maximizing the Q-value, the system can data-drivenally match different strategies to the most suitable scenarios: pure visual perception strategies typically achieve higher Q-values ​​in simple scenarios such as "structured, high visibility"; MSF tends to exhibit the highest Q-value in scenarios with "high occlusion and complex interactions"; while Transformer-based fusion strategies show superior Q-values ​​in scenarios such as "high-density traffic". The optimal strategy a is determined... * Then, the system will trigger the strategy preloading module to load the perception model and related configuration files required for the optimal strategy into the active memory of the computing unit in advance. When the vehicle enters the small area, the adaptive switching control module will activate the strategy.

[0084] In step S4, the vehicle locates its current position in real time while it is moving. Upon entering a new inter-cell area, the perception model and parameter configuration matching the scenario type are automatically switched and activated, specifically including:

[0085] During actual vehicle operation, the onboard high-precision positioning module continuously outputs the vehicle's precise geographical coordinates and heading information by fusing the matching results of the Global Navigation Satellite System and high-precision maps. The adaptive handover control module continuously compares this real-time location with the perception strategy map. When the system detects that the vehicle's current location has crossed the boundary of the next cell, a handover event is triggered. At this time, the adaptive handover control module immediately sends a control command containing a new strategy identifier to the adaptive perception execution module. Upon receiving this command, the perception system execution unit immediately activates the perception model and corresponding parameter configuration that have been pre-loaded into memory by the strategy pre-loading module and match the scene type between the new cell. Simultaneously, the system smoothly unloads or suspends the perception model of the previous cell to release computing resources.

[0086] In step S5, data such as scene features, strategy selection, and switching results are recorded in local and cloud databases to continuously optimize the parameters of the scene classification and strategy mapping model. Specifically, this includes:

[0087] After the strategy is activated and executed for a period of time, the vehicle-mounted performance evaluation and reward generation module will be responsible for quantitatively evaluating the actual performance of the current perception strategy and generating an instant reward value r, the specific expression of which can be represented as:

[0088] r = w perf ·(F1 currebt -F1 avg (s))-(w lat Latency norm +w gpu GPU norm )

[0089] Among them, w perf w lat w gpu Latency is a preset weighting factor. norm and GPU norm These represent the normalized model computation latency and GPU utilization, respectively, for F1. avg (s) represents the average performance in this scenario obtained through comparative test experiments when the initial Q value is established, F1. current This represents the actual performance of the current perception strategy.

[0090] Then, the data recording and uploading module packages the complete information constituting this decision-making and execution experience into an experience tuple (s, a, r, s') and uploads it to the cloud-based policy generation and optimization platform. Here, s represents the scenario category before policy execution, a represents the actual perceived policy executed, r represents the calculated immediate reward, and s' represents the new scenario category the vehicle is currently in. The reinforcement learning policy optimization module uses this data to continuously optimize the Q-value stored in the "scenario-optimal policy" mapping database. Its specific expression can be represented as:

[0091] Q(s,a)==Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)]

[0092] Where Q(s,a) is the current Q value to be updated, α is the parameter controlling the update step size, r is the immediate reward obtained in this experience, γ is the discount factor that weighs immediate and future rewards, and max a′ Q(s′,a′) is the maximum expected Q value that can be obtained in the new state s′, as retrieved from the database.

[0093] In summary, this invention provides an adaptive perception strategy switching method and system for long-distance autonomous driving. The invention presents the system's general architecture and workflow. Based on multi-objective joint optimization clustering technology, and combined with an initial mapping table constructed from large-scale experimental data and a reinforcement learning optimization mechanism, this invention presents the first detailed construction of an autonomous driving perception strategy switching scheme that integrates prior knowledge and online adaptive learning. Compared with existing single-perception schemes, the adaptive switching method proposed in this invention can intelligently select and activate the optimal perception strategy in different scenarios. While ensuring safety, it significantly improves the system's operating efficiency and environmental adaptability, providing a theoretical foundation and technical support for achieving safer, more efficient, and more intelligent high-order autonomous driving.

[0094] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An adaptive perception strategy switching method for long-distance autonomous driving, characterized in that, Includes the following steps: S1: The public dataset is analyzed using a multi-objective joint optimization clustering method to generate standard scenario prototypes, and then an initial scenario-optimal strategy mapping database is constructed. S2: Obtain the environmental, traffic, and dynamic target multi-dimensional scene features of the complete planned navigation route, perform global optimal scene division and identification of the target road segment, and generate a perception strategy map containing a series of continuous cells with scene labels. S3: Query the scene-optimal policy mapping database continuously optimized by reinforcement learning, select the optimal perception policy for each interval based on the Q-value maximization principle, and trigger the preloading module; S4: By monitoring the vehicle's position, the adaptive switching control module automatically activates the pre-loaded optimal perception strategy when the vehicle enters a new small area. S5: Evaluate the effectiveness of strategy execution and generate immediate rewards. Upload experience tuples containing status, actions, and rewards to the cloud and continuously optimize the Q-value database used through a temporal difference learning algorithm.

2. The adaptive perception strategy switching method for long-distance autonomous driving according to claim 1, characterized in that, In step S1, a multi-objective joint optimization clustering method is first used on the cloud server to generate a set of statistically optimal standard scene categories from scene fragments extracted from the public dataset, and an initial knowledge base is built based on these categories. The core objective is to find an optimal set C of clustering schemes. * With feature subset F * Its expression is: (C * ,F * )=argmin C,F J(C,F), J(C,F)=αD intra (C)+βO inter (C)+γR feat (F) Among them, D intra (C) represents the intra-cluster dispersion, O inter (C) represents the inter-cluster overlap, R feat (F) represents feature effectiveness, where α, β, and γ are adjustable hyperparameters used to balance the importance of different optimization objectives; A genetic algorithm is used to search for an approximate optimal solution to the above optimization problem. First, a clustering scheme (C,F) is encoded into a chromosome G. The chromosome contains an integer vector of the cluster label to which each data point belongs and a binary mask of the selected feature subset. Then, a clustering scheme containing N is randomly generated. pop Initial populations of different chromosomes For each chromosome G in the population j Calculate its corresponding fitness score, the formula is as follows: Among them, J(C j ,F j ) represents the joint optimization function value corresponding to the decoded chromosome, and ε is a very small normal number; The population is optimized by repeatedly performing a series of evolutionary operations; in each generation, a roulette wheel selection strategy is first used, based on the fitness score (G). j The chromosome is probabilistically selected from the current population for reproduction; the probability p of each chromosome being selected is... j It is proportional to its fitness, and its expression is: Then, with a preset crossover probability p c The selected parents are paired and their gene segments are exchanged to generate new offspring chromosomes; finally, with a mutation probability p... m The newly generated offspring chromosomes are randomly perturbed to maintain population diversity; the iterative evolution process is repeated until the preset maximum number of iterations T is reached. max The algorithm terminates when the fitness of the best individual in the population no longer significantly improves over multiple generations; at this point, the chromosome with the highest fitness in the current population is considered the final approximate optimal solution (Ci). * ,F * ); Each final cluster is defined as a standard scene prototype, forming a standard scene prototype library. Based on the results of systematic comparative testing experiments on various perception strategies under these standard scene categories, an initial Q value Q(s,a) is assigned to each scene prototype (state s) and each available perception strategy (action a). All initial Q values ​​together constitute the initial scene-optimal strategy mapping database.

3. The adaptive perception strategy switching method for long-distance autonomous driving according to claim 2, characterized in that, Intra-cluster dispersion D intra (C) is used to quantify the discreteness of scene features within all clusters, with the goal of ensuring high homogeneity within each scene category; Its expression is: Where C = {c1, c2, ..., c} k Let x be a scheme that divides all scene fragments into K clusters. i Belongs to cluster c k A scene feature vector, while μ k Then w is the centroid of the cluster. i To compare with sample x i The relevant sample weights, W, are used to characterize the representativeness or confidence level of the sample. k This is a feature weighting matrix used to define the importance of different feature dimensions when calculating intra-cluster dispersion.

4. The adaptive perception strategy switching method for long-distance autonomous driving according to claim 2, characterized in that, Inter-cluster overlap O inter (C) is used to quantify the boundary ambiguity between different clusters, with the goal of maximizing the distinguishability between different scene categories; Its expression is: Where, p k (x) and p j (x) represent clusters c k and c j The probability density function is given by μ, where μ and Σ are the mean vector and covariance matrix of the cluster, respectively.

5. The adaptive perception strategy switching method for long-distance autonomous driving according to claim 2, characterized in that, Feature effectiveness R feat (F) is used to quantify the information redundancy of the currently selected feature subset F when expressing scene differences. Its goal is to obtain a feature combination with a large amount of information but low internal correlation. Its expression is: Where F is the feature subset used in the current scheme; |F| is the size of the subset; double summation ΣΣ represents calculating the sum of mutual information between all distinct feature pairs in the subset, I N (f i ;f j ) is a feature f i and feature f j Normalized mutual information between them.

6. The adaptive perception strategy switching method for long-distance autonomous driving according to claim 1, characterized in that, In step S2, firstly, for the predetermined long-distance target driving route, multi-source data is fused to construct an instantaneous, multi-dimensional scene feature stream; then, the continuous scene feature stream is scanned, and any feature that undergoes a significant change is marked as a potential change point, generating a set P = {p1, p2, ..., p...} of N ordered candidate segmentation points. N }; then for any two candidate points p j and p i For each defined potential subinterval seg(j,i), a unique interval matching cost is calculated, and the best standard scene cluster it matches is recorded, 1≤j≤i≤N; a set of M standard scene clusters C={c1,c2,...,c M In the library, the standard scene cluster that is closest to μ(seg(j,i)) is found through nearest neighbor matching. Calculate the distance from all points within the potential interval to the centroid of this optimal standard cluster. The average squared distance is used as its matching cost C. match (seg(j,i)); Its expression is: Where, μ m For standard scene cluster c m The centroid of point p; F(p) represents the normalized feature vector of point p. All calculated matching costs are stored in a two-dimensional matrix MatchCostMatrix[j][i]. At the same time, the best standard cluster index m* matched by each potential interval seg(j,i) is also recorded in another matrix BestClusterIndex[j][i]. Then, a dynamic programming algorithm is used to find the optimal K-1 final split points among the N candidate split points; a two-dimensional table Cost[i][k] is constructed to store the minimum total cost of dividing the road segment covered by the first i candidate points into k cells; this two-dimensional table is filled with the following recursive relation: Among them, C seg (j,i) represents the candidate point p. j to p i The total cost generated by this segment of the road as a single small interval; this total cost consists of its own matching cost and the switching cost with the previous interval, and its expression is: C seg (j,i)=MatchCostMatrix[j+1][i]+λ·C switch (with j-1 ,seg(j+1,i)) Among them, the switching cost C switch The cost of switching from the scenario type of the previous optimal interval to the scenario type of the current new evaluation interval is expressed as follows: Where ε is a very small positive constant to avoid the denominator being zero; and To obtain the previous optimal interval s by querying the BestClusterIndex matrix. k-1 The best standard cluster index that matches each of the current new intervals seg(j+1,i); After the dynamic programming algorithm has finished executing and filled the entire Cost table, the optimal segmentation scheme for the specific route can be obtained by backtracking Cost[N][K]. This series of small intervals and labels constitutes the perception strategy map for this long-distance task.

7. The adaptive perception strategy switching method for long-distance autonomous driving according to claim 1, characterized in that, In step S3, based on the pre-built scenario-optimal strategy mapping database and a data-driven intelligent decision-making model, the optimal perception strategy is matched and selected for each cell, and the corresponding optimal perception strategy is pre-loaded into the vehicle's local system. Specifically, this includes: When selecting a policy between cells, the scenario-optimal policy mapping module queries the scenario-optimal policy mapping database, which is continuously optimized by the reinforcement learning policy optimization module. For a standard scenario type (state s) that has been classified in step S2, the intelligent decision model selects from the set A = {a1, a2, ..., a...} of all available perception policies corresponding to that scenario. n In the given information, select strategy a that has the highest expected Q value. * The expression for this decision-making process is: a * =argmax a∈A Q(s,a) Among them, a * The selected optimal strategy is represented by A; A represents the set of all available perception strategies in the system, mainly including pure visual perception strategies, multi-sensor fusion strategies, and Transformer-based fusion strategies; Q(s,a) is a quantitative estimate of the expected long-term reward of performing action a in state s, retrieved from the database; when determining the optimal strategy a... * Then, the strategy preloading module is triggered to preload the perception model and related configuration files required for the optimal strategy into the active memory of the computing unit. When the vehicle enters the small area, the adaptive switching control module activates the strategy.

8. The adaptive perception strategy switching method for long-distance autonomous driving according to claim 1, characterized in that, In step S4, during vehicle operation, the vehicle continuously locates its current position; upon entering a new inter-cell area, it automatically switches and activates a perception model and parameter configuration matching the scenario type, specifically including: During actual vehicle operation, the onboard high-precision positioning module continuously outputs the vehicle's precise geographical coordinates and heading information by fusing the matching results of the Global Navigation Satellite System and high-precision maps. The adaptive switching control module continuously compares this real-time location with the perception strategy map. When the system detects that the vehicle's current location has crossed the boundary of the next cell, a switching event is triggered. At this time, the adaptive switching control module immediately sends a control command containing a new strategy identifier to the adaptive perception execution module. Upon receiving the command, the perception system execution unit immediately activates the perception model and corresponding parameter configuration that are pre-loaded into memory by the strategy pre-loading module and match the scene type between the new cells. Simultaneously, the system unloads or suspends the perception model of the previous cell to release computing resources.

9. The adaptive perception strategy switching method for long-distance autonomous driving according to claim 1, characterized in that, In step S5, scene features, strategy selection, and switching result data are recorded in local and cloud databases to continuously optimize scene classification and strategy mapping model parameters, specifically including: After the strategy is activated and executed for a period of time, the vehicle-mounted performance evaluation and reward generation module is responsible for quantitatively evaluating the actual performance of the current perception strategy and generating an instant reward value r, the expression of which is: r=w perf ·(F1 current -F1 avg (s))-(w lat ·Latency norm +w gpu ·GPU norm ) Among them, w perf w lat w gpu Latency is a preset weighting factor. norm and GPU norm These represent the normalized model computation latency and GPU utilization, respectively, for F1. avg (s) represents the average performance in this scenario obtained through comparative test experiments when the initial Q value is established, F1. current This represents the actual performance of the current perception strategy. Then, the data recording and uploading module packages the complete information constituting this decision-making and execution experience into an experience tuple (s, a, r, s') and uploads it to the cloud-based policy generation and optimization platform, where s is the scenario category before the policy is executed, a is the actual perceived policy executed, r is the just-calculated immediate reward, and s' is the new scenario category where the vehicle is currently located. The reinforcement learning policy optimization module uses this data to continuously optimize the Q-value stored in the scenario-optimal policy mapping database; its expression is: Q(s,a)==Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)] Where Q(s,a) is the current Q value to be updated, α is the parameter controlling the update step size, r is the immediate reward, γ is the discount factor that weighs immediate and future rewards, and max a′ Q(s′,a′) is the maximum expected Q value that can be obtained in the new state s′, as retrieved from the database.

10. An adaptive perception strategy switching system for long-distance autonomous driving, characterized in that, It includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it can implement the method as described in any one of claims 1-9.

Images