Automatic driving system extreme environment decision-making method based on neural evolutionary game

Abstract

The invention discloses an automatic driving system extreme environment decision-making method based on a neural evolutionary game. The method comprises the following steps: S1, collecting and constructing a real-time environment data set around a vehicle through an automatic driving system; s2, generating an environment feature vector; s3, based on the environment feature vector, analyzing the environment feature data through an extreme environment identification module; s4, activating a corresponding multi-target dynamic reward mechanism based on the extreme environment label, and setting a dynamic reward weight for the neural evolution process; s5, gradually evolving to generate an optimal neural network structure and parameter configuration suitable for different extreme environments; s6, generating a multi-target game strategy in combination with the environment feature vector; s7, generating an optimal driving path, vehicle speed adjustment, obstacle avoidance mode and direction control decision; and S8, inputting real-time feedback data into the neural evolution model, and further adjusting the structure and parameter configuration of the neural network. The automatic driving system can make a safer, more accurate and more efficient driving decision in an extreme environment.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an extreme environment decision-making method for an autonomous driving system based on neuroevolutionary game. Background Art

[0002] With the development of autonomous driving technology, the ability of vehicles to make autonomous decisions in different environments has become the key to improving autonomous driving safety and driving experience. The decision-making process of autonomous driving systems faces severe challenges in complex and extreme environments. Existing autonomous driving decision-making methods mainly rely on sensor data and environmental models, and perform environmental perception and decision-making through predefined rules or based on traditional machine learning methods. Although they have achieved certain success in some conventional driving scenarios, their performance and adaptability are significantly insufficient in extreme environments.

[0003] A common method in the prior art is an autonomous driving decision-making system based on perception algorithms and simple rules. The autonomous driving decision-making system usually uses traditional sensor input for environmental perception, and then makes driving decisions based on fixed rules or models. It can effectively complete the task under normal conditions, but the system often faces greater limitations in complex and extreme environments. The recognition accuracy of sensors in low-visibility environments such as fog, heavy rain or snow will be seriously affected, resulting in inaccurate environmental perception. At the same time, rule-based decision-making methods are difficult to make timely and flexible responses in dynamically changing complex environments, which can easily lead to decision-making delays or unsafe behaviors.

[0004] Another type of autonomous driving decision-making method is based on deep learning and neural network models, which learn driving strategies in complex environments through large amounts of data training. Although neural networks can improve decision-making accuracy to a certain extent, most of the existing neural network models are based on static training and lack dynamic adjustment mechanisms. When vehicles enter different extreme environments, traditional neural network models often cannot effectively adjust their strategies to cope with new environmental changes. Although they can provide more accurate decisions in specific situations, they lack adaptability to complex and changing environments and real-time optimization capabilities, and their perception and decision-making in extreme environments are still not flexible and reliable enough.

[0005] Therefore, the shortcomings of existing autonomous driving technology mainly include: first, traditional perception and decision-making methods have poor adaptability in extreme environments and cannot effectively cope with complex environmental changes; second, existing deep learning-based decision-making methods lack dynamic adjustment mechanisms for extreme environments and cannot optimize decision-making strategies in real time; finally, existing technologies ignore the modeling of dynamic interactive relationships among multiple parties and lack effective prediction and comprehensive optimization of the behaviors of vehicles, pedestrians and obstacles. Summary of the invention

[0006] One purpose of the present invention is to propose an extreme environment decision-making method for an autonomous driving system based on neuroevolutionary game. The present invention enables the autonomous driving system to make safer, more accurate and efficient driving decisions in extreme environments.

[0007] According to an embodiment of the present invention, a method for making decisions in extreme environments of an autonomous driving system based on neuroevolutionary game includes the following steps:

[0008] S1. Collect and construct a real-time environmental data set around the vehicle through the autonomous driving system to obtain weather data, road condition data, and obstacle location data;

[0009] S2. Input the real-time environmental data set into the data processing unit for data cleaning, noise filtering and feature extraction to obtain environmental feature data, and extract potential obstacle information of different risk levels according to the environmental feature data to generate an environmental feature vector;

[0010] S3. Based on the environmental feature vector, the environmental feature data is analyzed by the extreme environment recognition module to determine whether the environment in which the vehicle is located belongs to a predefined extreme environment type, which includes rainy and snowy weather, foggy weather, low visibility and complex mountain road conditions; if it is identified as an extreme environment, the extreme environment label is output to the neural evolution module to trigger different multi-objective dynamic reward mechanisms;

[0011] S4. Activate a multi-objective dynamic reward mechanism corresponding to the extreme environment label, and set a dynamic reward weight for the neural evolution process. The reward weight sets different decision priorities according to different extreme environment types to prioritize strengthening the optimal adaptive strategy in the specific extreme environment.

[0012] S5. Under the guidance of the multi-objective dynamic reward mechanism, the decision optimization model is selected, mutated and reorganized through the neuroevolution model, and the optimal neural network structure and parameter configuration that can adapt to different extreme environments are gradually evolved;

[0013] S6. Based on the optimal neural network structure, the game theory model is introduced to build a multi-factor comprehensive optimization decision-making model. The game theory model establishes a multi-party behavior prediction model in a complex and extreme environment by simulating the interactive relationship between vehicles, pedestrians and dynamic obstacles, and generates a multi-objective game strategy in combination with the environmental feature vector;

[0014] S7, using a multi-objective game strategy, which generates the optimal driving path, speed adjustment, obstacle avoidance method and direction control decision based on real-time environmental data and predicted behavior;

[0015] S8. During the actual driving process, the system inputs the multi-objective game strategy execution feedback data, including vehicle execution results, environmental changes and potential new risk information, into the neural evolution model through real-time feedback data to further adjust the structure and parameter configuration of the neural network.

[0016] Optionally, the S1 step specifically includes:

[0017] S11. Using the visual camera in the autonomous driving system to collect real-time image data for detecting visual information in the environment v , using the laser radar in the autonomous driving system to collect point cloud data for generating obstacle information in three-dimensional space l , using millimeter-wave radar in the autonomous driving system to collect target distance and speed information for identifying dynamic objects m , using ultrasonic sensors in the autonomous driving system to collect close-range obstacle data for detecting obstacles at low speed or in a stationary state. u , all sensor data form the initial environment dataset

[0018] S12. Initial environment data set Perform time series calibration and synchronize the data with the sampling time t to obtain the time-aligned environmental data sequence D env (t);

[0019] S13, based on environmental data sequence D env (t) data format, decomposing the data into three types of feature information: weather data D weather (t), road condition data D road (t) and obstacle position data D obs (t) and build a real-time environment dataset

[0020]

[0021] Among them, f is the multimodal data fusion and feature extraction function.

[0022] Optionally, the S2 step specifically includes:

[0023] S21. Real-time environment dataset Input to the data processing unit, perform data cleaning, and generate the cleaned environmental data set D by removing outliers and correcting sampling errors. clean (t);

[0024] S22, the cleaned environmental data set D clean(t) Perform noise filtering and use a multimodal data joint noise filtering algorithm to process the specific noise characteristics of weather data, road condition data, and obstacle location data, respectively, to generate a noise-filtered environmental data set D filtered (t);

[0025] S23, environmental data set D after noise filtering filtered (t) Perform feature extraction and extract weather features F through feature extraction algorithm weather (t), road condition characteristics F road (t) and obstacle characteristics F obs (t), generate the environmental feature vector V env (t):

[0026] V env (t) = {F weather (t),F road (t),F obs (t)}.

[0027] Optionally, the S3 step specifically includes:

[0028] S31, the environmental feature vector V env (t) Input to the extreme environment recognition module, analyze the weather characteristics, road conditions and obstacle characteristics, and for each predefined extreme environment type E i Calculate each characteristic and extreme environment type E i Matching score:

[0029] Calculate weather feature matching score

[0030]

[0031] in, is the kth weather characteristic index collected at time t, which affects the perception and decision-making ability of the autonomous driving system under extreme weather conditions. In order to map the kth weather characteristic index into a feature mapping function for the matching score of the extreme environment type, the mapping function outputs a high matching score when the extreme environment type is foggy weather and the visibility is low. is the kth weather characteristic index in the extreme environment type E i The importance weight in

[0032] Calculate the road condition feature matching score

[0033]

[0034] in, is the lth road condition characteristic index collected at time t, which affects the driving stability and handling performance of the vehicle. In order to map the lth road condition characteristic index into a characteristic mapping function for the matching score of the extreme environment type, a high matching score is output for the road slope characteristic of the complex mountainous road condition when the slope is steep. is the lth road condition characteristic index in extreme environment type E i The importance weight in

[0035] Calculate obstacle feature matching score

[0036]

[0037] in, is the mth obstacle characteristic index collected at time t, which affects the vehicle’s obstacle avoidance strategy and safety. In order to map the mth obstacle feature index into a feature mapping function for the matching score of the extreme environment type, the obstacle density feature of the low visibility environment outputs a high matching score when the obstacles are dense. is the characteristic index of the mth obstacle in the extreme environment type E i The importance weight in

[0038] S32. Calculate the extreme environment type E based on the matching scores of each feature i The comprehensive matching score M i( t ) :

[0039]

[0040] Among them, α i , β i , γ i For extreme environment type E i The global weight coefficient under reflects the importance of weather characteristics, road conditions and obstacle characteristics in identifying extreme environment types;

[0041] S33, the comprehensive matching score M i( t ) Compared with the preset extreme environment recognition threshold T i For comparison, if M i (t)≥T i , then the current environment is judged to be an extreme environment type E i , generate the extreme environment label L(t):

[0042] L(t)={E i ∣M i (t)≥T i};

[0043] Where L(t) is the set of extreme environment types identified at time t, including all extreme environment types that meet the conditions, T i It is the identification threshold of extreme environment types, which is set in advance based on experiments and experience.

[0044] Optionally, the S4 step specifically includes:

[0045] S41. Activate the corresponding multi-objective dynamic reward mechanism according to the extreme environment label L(t) identified in step S3, and set the dynamic reward weight W for the neural evolution process. i reward (t), where i represents the i-th extreme environment type;

[0046] S42, in rainy and snowy weather conditions E rain ∈L(t), setting the reward weights for stable driving and braking distance For high:

[0047]

[0048] Where v(t) is the actual speed of the vehicle at time t, v safe The safe driving speed in rainy and snowy weather is dynamically adjusted according to the slippery degree of the road and the intensity of rainfall. v is the speed deviation tolerance, indicating the permissible range of speed stability, d brake( t ) is the actual braking distance of the vehicle at time t, d ideal is the ideal braking distance in rainy and snowy weather, σ d is the tolerance of braking distance deviation, η1 and η2 are reward weight coefficients, which respectively reflect the importance of stable driving and braking distance in rainy and snowy weather;

[0049] S43, in foggy weather conditions E fog ∈L ( t ) , set the reward weight for obstacle avoidance ability For high:

[0050]

[0051] Among them, d obs( t ) is the distance between the vehicle and the obstacle in front at time t, d critical is the critical obstacle avoidance distance in foggy weather, λ1 is the distance sensitivity coefficient, which controls the rate at which the obstacle avoidance reward changes with distance, and t response is the obstacle avoidance reaction time of the vehicle at time t, reflecting the agility of the system, t optimal is the optimal obstacle avoidance reaction time in foggy weather, σt is the reaction time deviation tolerance, which indicates the obstacle avoidance reaction speed requirement, θ1 and θ2 are the reward weight coefficients;

[0052] S44. In low-visibility nighttime conditions E night ∈L ( t ) , set the reward weights for object recognition accuracy and reaction speed For high:

[0053]

[0054] Among them, N correct( t ) is the number of objects correctly identified by the vehicle at time t, N total( t ) is the total number of objects actually existing in the environment, Represents the exponential weighting of object recognition accuracy, γ1 is the adjustment index, t decision is the time when the vehicle makes a driving decision at time t, t ideal is the ideal decision response time in nighttime environment, σ τ is the tolerance of decision time deviation, κ1 and κ2 are reward weight coefficients, reflecting the importance of object recognition accuracy and reaction speed in low visibility environment;

[0055] S45, in complex mountainous road conditions E mountain ∈L ( t ) , giving priority to the reward weights of real-time path planning and obstacle avoidance For high:

[0056]

[0057] Among them, C path ( t ) is the sum of the curvature change rates of the vehicle's planned path at time t, δ1 is the path complexity sensitivity coefficient, N collision ( t ) The number of collisions that occur during the vehicle's obstacle avoidance process at time t, N attempts ( t ) is the total number of times the vehicle attempts to avoid obstacles, represents the exponential weighting of the obstacle avoidance success rate, β1 is the adjustment index, ξ1 and ξ2 are the reward weight coefficients, which respectively reflect the importance of path planning and obstacle avoidance in complex mountainous road conditions.

[0058] Optionally, the step S6 specifically includes:

[0059] S61, based on the optimal neural network structure N(t) and the environmental feature vector V generated in step S5 env( t ) , combining the extreme environment label L(t) and the state of dynamic objects, a multi-party game environment model is established, where the dynamic objects include vehicles A, pedestrians P and other dynamic obstacles O;

[0060] S62, construct the state space S(t), action space A(t) and benefit function U(t) of the game model;

[0061] S63. Establish a behavior prediction model based on the multi-party game environment model to calculate the optimal strategy π of each game participant A ( t ) ,π P( t ) ,π O( t ) :

[0062]

[0063] in, are the optimal strategies for vehicle A, pedestrian P and obstacle O, respectively, P( t and U O( t ) is the profit function of pedestrians and obstacles;

[0064] S64. Output the optimal strategy generated by the game model to the decision-making module, and generate multi-objective optimized driving decisions in combination with real-time environmental feature data, including speed adjustment, direction control and path selection.

[0065] The beneficial effects of the present invention are:

[0066] (1) The present invention introduces a dynamic reward mechanism based on extreme environment labels. The dynamic reward mechanism can adaptively adjust the reward weight according to different extreme environment types (rainy and snowy weather, heavy fog, night, and complex mountain road conditions). By targetedly strengthening the driving strategy under specific environmental conditions, the present invention can significantly improve the decision-making accuracy and flexibility of the autonomous driving system in complex environments. In rainy and snowy weather, the system sets a higher reward weight for stable driving and braking distance. In foggy weather, the system prioritizes the obstacle avoidance capability, thereby effectively improving the environmental adaptability and safety of the system. Compared with traditional decision-making methods based on fixed rules or simple machine learning models, the dynamic reward mechanism of the present invention can adjust the decision-making strategy in real time according to environmental changes, ensuring that the autonomous driving system can make safer, more accurate and efficient driving decisions in extreme environments.

[0067] (2) The present invention constructs a multi-factor comprehensive optimization decision model by combining the game theory model to simulate the interaction between vehicles, pedestrians and dynamic obstacles. The game theory model can not only handle complex dynamic interaction behaviors, but also generate multi-objective game strategies based on environmental feature vectors, realizing accurate prediction and optimization of the interaction relationship between multiple actors. In traditional autonomous driving systems, the complexity of these multi-party dynamic interactions is often ignored, resulting in inaccurate or risky decisions in dynamic environments. The introduction of the game theory model can effectively solve this problem, so that the vehicle can make more accurate predictions and decisions based on real-time environmental changes when facing multi-party dynamic behaviors, greatly improving the intelligence level and reliability of the autonomous driving system in complex scenarios.

[0068] (3) The present invention adopts a neuroevolutionary game algorithm, and introduces a multi-objective optimized neural network structure to adaptively adjust driving decisions in extreme environments. Different from traditional deep learning methods, the present invention introduces a multi-objective dynamic reward mechanism in the neuroevolution process, and performs real-time weight adjustment in combination with environmental characteristics, thereby optimizing the structure and decision-making ability of the neural network, and being able to dynamically adjust the decision-making strategy according to environmental changes, so that the system can not only quickly adapt and make accurate decisions when dealing with complex and changeable extreme environments, but also effectively avoid the overfitting or underfitting problems of traditional models when the environment changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 This is a flowchart of an extreme environment decision-making method for an autonomous driving system based on neuroevolutionary game proposed by the present invention. DETAILED DESCRIPTION

[0071] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0072] refer to Figure 1 , a decision-making method for an autonomous driving system in extreme environments based on neuroevolutionary games, comprising the following steps:

[0073] S1. Collect and construct a real-time environmental data set around the vehicle through the autonomous driving system to obtain weather data, road condition data, and obstacle location data;

[0074] S2. Input the real-time environmental data set into the data processing unit for data cleaning, noise filtering and feature extraction to obtain environmental feature data, and extract potential obstacle information of different risk levels according to the environmental feature data to generate an environmental feature vector;

[0075] S3. Based on the environmental feature vector, the extreme environment recognition module analyzes the environmental feature data to determine whether the vehicle's environment belongs to a predefined extreme environment type, which includes rainy and snowy weather, foggy weather, low visibility, and complex mountain road conditions. If it is identified as an extreme environment, the extreme environment label is output to the neural evolution module to trigger different multi-objective dynamic reward mechanisms.

[0076] S4. Based on the multi-objective dynamic reward mechanism corresponding to the activation of extreme environment labels, dynamic reward weights are set for the neural evolution process. The reward weights set different decision priorities according to different extreme environment types to prioritize the optimal adaptive strategy in this specific extreme environment.

[0077] S5. Under the guidance of the multi-objective dynamic reward mechanism, the decision optimization model is selected, mutated and reorganized through the neuroevolution model, and the optimal neural network structure and parameter configuration that can adapt to different extreme environments are gradually evolved;

[0078] S6. Based on the optimal neural network structure, the game theory model is introduced to build a multi-factor comprehensive optimization decision-making model. The game theory model establishes a multi-party behavior prediction model in a complex and extreme environment by simulating the interactive relationship between vehicles, pedestrians and dynamic obstacles, and generates a multi-objective game strategy in combination with the environmental feature vector;

[0079] S7, using a multi-objective game strategy, which generates the optimal driving path, speed adjustment, obstacle avoidance method and direction control decision based on real-time environmental data and predicted behavior;

[0080] S8. During the actual driving process, the system inputs the multi-objective game strategy execution feedback data, including vehicle execution results, environmental changes and potential new risk information, into the neural evolution model through real-time feedback data to further adjust the structure and parameter configuration of the neural network.

[0081] In this implementation, step S1 specifically includes:

[0082] S11. Using the visual camera in the autonomous driving system to collect real-time image data for detecting visual information in the environment v , using the laser radar in the autonomous driving system to collect point cloud data for generating obstacle information in three-dimensional space l , using millimeter-wave radar in the autonomous driving system to collect target distance and speed information for identifying dynamic objects m, using ultrasonic sensors in the autonomous driving system to collect close-range obstacle data for detecting obstacles at low speed or in a stationary state. u , all sensor data form the initial environment dataset

[0083] S12. Initial environment data set Perform time series calibration and synchronize the data with the sampling time t to obtain the time-aligned environmental data sequence D env (t);

[0084] S13, based on environmental data sequence D env (t) data format, decomposing the data into three types of feature information: weather data D weather (t), road condition data D road (t) and obstacle position data D obs (t) and build a real-time environment dataset

[0085]

[0086] Among them, f is the multimodal data fusion and feature extraction function.

[0087] In this implementation, step S2 specifically includes:

[0088] S21. Real-time environment dataset Input to the data processing unit, perform data cleaning, and generate the cleaned environmental data set D by removing outliers and correcting sampling errors. clean (t);

[0089] S22, the cleaned environmental data set D clean (t) Perform noise filtering and use a multimodal data joint noise filtering algorithm to process the specific noise characteristics of weather data, road condition data, and obstacle location data, respectively, to generate a noise-filtered environmental data set D filtered (t);

[0090] S23, environmental data set D after noise filtering filtered (t) Perform feature extraction and extract weather features F through feature extraction algorithm weather (t), road condition characteristics F road (t) and obstacle characteristics F obs (t), generate the environmental feature vector V env (t):

[0091] V env (t) = {F weather (t),F road (t),Fobs (t)}.

[0092] In this implementation, step S3 specifically includes:

[0093] S31, the environmental feature vector V env (t) Input to the extreme environment recognition module, analyze the weather characteristics, road conditions and obstacle characteristics, and for each predefined extreme environment type E i Calculate each characteristic and extreme environment type E i Matching score:

[0094] Calculate weather feature matching score

[0095]

[0096] in, is the kth weather characteristic index collected at time t, which affects the perception and decision-making ability of the autonomous driving system under extreme weather conditions. In order to map the kth weather characteristic index into a feature mapping function for the matching score of the extreme environment type, the mapping function outputs a high matching score when the extreme environment type is foggy weather and the visibility is low. is the kth weather characteristic index in the extreme environment type E i The importance weight in

[0097] Calculate the road condition feature matching score

[0098]

[0099] in, is the lth road condition characteristic index collected at time t, which affects the driving stability and handling performance of the vehicle. In order to map the lth road condition characteristic index into a characteristic mapping function for the matching score of the extreme environment type, a high matching score is output for the road slope characteristic of the complex mountainous road condition when the slope is steep. is the lth road condition characteristic index in extreme environment type E i The importance weight in

[0100] Calculate obstacle feature matching score

[0101]

[0102] in, is the mth obstacle characteristic index collected at time t, which affects the vehicle’s obstacle avoidance strategy and safety. In order to map the mth obstacle feature index into a feature mapping function for the matching score of the extreme environment type, the obstacle density feature of the low visibility environment outputs a high matching score when the obstacles are dense. is the characteristic index of the mth obstacle in the extreme environment type E i The importance weight in

[0103] S32. Calculate the extreme environment type E based on the matching scores of each feature i The comprehensive matching score M i( t ) :

[0104]

[0105] Among them, α i , β i , γ i For extreme environment type E i The global weight coefficient under reflects the importance of weather characteristics, road conditions and obstacle characteristics in identifying extreme environment types;

[0106] S33, the comprehensive matching score M i( t ) Compared with the preset extreme environment recognition threshold T i For comparison, if M i( t ) ≥T i , then the current environment is judged to be an extreme environment type E i , generate the extreme environment label L(t):

[0107] L ( t ) = { E i ∣M i ( t ) ≥T i } ;

[0108] Where L(t) is the set of extreme environment types identified at time t, including all extreme environment types that meet the conditions, T i It is the identification threshold of extreme environment types, which is set in advance based on experiments and experience.

[0109] In this implementation, step S4 specifically includes:

[0110] S41. Activate the corresponding multi-objective dynamic reward mechanism according to the extreme environment label L(t) identified in step S3, and set the dynamic reward weight for the neural evolution process. Where i represents the i-th extreme environment type;

[0111] S42, in rainy and snowy weather conditions E rain ∈L ( t ) , set the reward weights for stable driving and braking distance For high:

[0112]

[0113] Where v(t) is the actual speed of the vehicle at time t, v safe The safe driving speed in rainy and snowy weather is dynamically adjusted according to the slippery degree of the road and the intensity of rainfall. v is the speed deviation tolerance, indicating the permissible range of speed stability, d brake( t ) is the actual braking distance of the vehicle at time t, d ideal is the ideal braking distance in rainy and snowy weather, σ d is the tolerance of braking distance deviation, η1 and η2 are reward weight coefficients, which respectively reflect the importance of stable driving and braking distance in rainy and snowy weather;

[0114] S43, in foggy weather conditions E fog ∈L ( t ) , set the reward weight for obstacle avoidance ability For high:

[0115]

[0116] Among them, d obs( t ) is the distance between the vehicle and the obstacle in front at time t, d critical is the critical obstacle avoidance distance in foggy weather, λ1 is the distance sensitivity coefficient, which controls the rate at which the obstacle avoidance reward changes with distance, and t response is the obstacle avoidance reaction time of the vehicle at time t, reflecting the agility of the system, t optimal is the optimal obstacle avoidance reaction time in foggy weather, σ t is the reaction time deviation tolerance, which indicates the obstacle avoidance reaction speed requirement, θ1 and θ2 are the reward weight coefficients;

[0117] S44. In low-visibility nighttime conditions E night ∈L ( t ) , set the reward weights for object recognition accuracy and reaction speed For high:

[0118]

[0119] Among them, N correct(t ) is the number of objects correctly identified by the vehicle at time t, N total( t ) is the total number of objects actually existing in the environment, Represents the exponential weighting of object recognition accuracy, γ1 is the adjustment index, t decision is the time when the vehicle makes a driving decision at time t, t ideal is the ideal decision response time in nighttime environment, σ τ is the tolerance of decision time deviation, κ1 and κ2 are reward weight coefficients, reflecting the importance of object recognition accuracy and reaction speed in low visibility environment;

[0120] S45, in complex mountainous road conditions E mountain ∈L ( t ) , giving priority to the reward weights of real-time path planning and obstacle avoidance For high:

[0121]

[0122] Among them, C path ( t ) is the sum of the curvature change rates of the vehicle's planned path at time t, δ1 is the path complexity sensitivity coefficient, N collision ( t ) The number of collisions that occur during the vehicle's obstacle avoidance process at time t, N attempts ( t ) is the total number of times the vehicle attempts to avoid obstacles, represents the exponential weighting of the obstacle avoidance success rate, β1 is the adjustment index, ξ1 and ξ2 are the reward weight coefficients, which respectively reflect the importance of path planning and obstacle avoidance in complex mountainous road conditions.

[0123] In this implementation, step S6 specifically includes:

[0124] S61, based on the optimal neural network structure N(t) and the environmental feature vector V generated in step S5 env( t ) , combining the extreme environment label L(t) and the state of dynamic objects, a multi-party game environment model is established, where the dynamic objects include vehicles A, pedestrians P and other dynamic obstacles O;

[0125] S62, construct the state space S(t), action space A(t) and benefit function U(t) of the game model;

[0126] S63. Establish a behavior prediction model based on the multi-party game environment model to calculate the optimal strategy π of each game participantA ( t ) ,π P( t ) ,π O( t ) :

[0127]

[0128] in, are the optimal strategies for vehicle A, pedestrian P and obstacle O, respectively, P( t and U O( t ) is the profit function of pedestrians and obstacles;

[0129] S64. Output the optimal strategy generated by the game model to the decision-making module, and generate multi-objective optimized driving decisions in combination with real-time environmental feature data, including speed adjustment, direction control and path selection.

[0130] Embodiment 1:

[0131] This embodiment 1 demonstrates the application process of an extreme environment decision-making method for an autonomous driving system based on neuroevolutionary game, and demonstrates the feasibility and effectiveness of the present invention by simulating actual driving scenarios in rainy, snowy and foggy weather.

[0132] Example 1 On December 15, 2024, the location was on the G104 National Highway in a certain mountainous area. The road section has high mountains and steep slopes all year round, and there is often fog, rain and snow in winter. Traffic accidents occur frequently in the area. Especially in extreme weather conditions, drivers face greater challenges. In order to reduce the accident rate and improve the safety and efficiency of the autonomous driving system, an autonomous driving vehicle equipped with the method based on the present invention is being tested.

[0133] The weather conditions at that time were alternating heavy snow and dense fog, with low visibility, strong wind speed, slippery roads, and a vehicle speed of 50 kilometers per hour. In addition to the self-driving vehicle, there was also an ordinary car and several pedestrians walking on the roadside. The purpose of the test was to verify how the self-driving system can make optimal decisions during dynamic driving according to different extreme environmental conditions to improve safety and driving efficiency.

[0134] During the entire driving process, the autonomous driving system adjusts driving decisions in real time based on its multi-objective decision-making mechanism based on the neuroevolutionary game model.

[0135] When the autonomous driving system enters the test area, it uses high-precision sensors (including lidar, millimeter-wave radar, and cameras) to obtain real-time data about the surrounding environment. Based on the sensor data and the vehicle's driving status, the system automatically identifies that the current environment is a mixed environment of rain, snow, and fog, and generates the corresponding environmental labels L(t), which are E snow and E fog .

[0136] The system then activated a decision-making model based on a multi-objective dynamic reward mechanism, in which the autonomous driving system set reward weights for rainy, snowy and foggy weather (in rainy and snowy weather, the reward weights for stable driving and braking distance are higher, while in foggy weather, the reward weights for obstacle avoidance ability are higher).

[0137] In rainy and snowy weather, the system first focuses on the vehicle's speed and braking distance. In the current situation, the vehicle speed is 50 kilometers per hour. According to the safe driving speed v in rainy and snowy weather safe = 40 kilometers per hour, the system automatically adjusts the vehicle's speed and dynamically calculates the stable driving reward weight through an algorithm:

[0138]

[0139] Substitute the actual parameters into the formula for calculation: v(t) = 50 kilometers per hour, vehicle speed v safe = 40 km / h, safe driving speed in rainy and snowy weather, d brake (t) = 25 meters, current braking distance, d ideal = 30 meters, ideal braking distance in rainy and snowy weather, σ v =5, speed deviation tolerance, σ d =2, braking distance deviation tolerance, η1=1.2, reward weight coefficient, η2=1.0, reward weight coefficient;

[0140] Calculation formula:

[0141]

[0142] First, calculate the speed part:

[0143]

[0144] Then calculate the braking distance:

[0145]

[0146] Therefore, the final calculation is:

[0147]

[0148] The system uses this reward weight to adjust the vehicle's speed and braking control in real time.

[0149] When the system detects foggy weather ahead, the reward mechanism prioritizes obstacle avoidance. At this time, the system obtains the distance d of the vehicle ahead through the camera and radar. obs( t ) = 50 meters, and the distance of the obstacle in front is d critical = 30 meters. According to the distance sensitivity coefficient λ = 10, and the obstacle avoidance reaction time t response = 1.5 seconds, the system calculates the obstacle avoidance reward weight;

[0150] Substitute the actual parameters into the formula for calculation: d obs( t ) = 50 meters, the distance of the obstacle in front, d critical =30 meters, critical obstacle avoidance distance, λ=10, distance sensitivity coefficient, t response = 1.5 seconds, actual obstacle avoidance reaction time, t optimal = 1.0 seconds, ideal reaction time, σ t =0.3, reaction time tolerance, θ1=1.5, reward weight coefficient, θ2=1.0, reward weight coefficient;

[0151] Calculation formula:

[0152]

[0153] First, calculate the obstacle avoidance distance:

[0154]

[0155] Then calculate the reaction time part:

[0156]

[0157] Therefore, the final calculation is:

[0158]

[0159] Through this reward weight, the system optimized the obstacle avoidance decision and responded 5 meters in advance to avoid collision with the obstacle ahead.

[0160] To verify the effectiveness of the present invention, we conducted a comparative test between the method of the present invention and a traditional autonomous driving system. In this test, the following data and scenarios were used.

[0161] Traditional method data: Vehicle speed: 50 km / h. When encountering rainy and snowy weather, the system failed to adjust the vehicle speed and braking distance in time. The braking distance was 35 meters, resulting in the need to decelerate in advance. When encountering foggy weather, the system failed to effectively predict the behavior of pedestrians in front. It only reacted when it was 2 meters away from the pedestrians, resulting in a reaction time of 2 seconds, avoiding collision time difference.

[0162] Data of the method of the present invention: Vehicle speed: 50 km / h. In rainy and snowy weather, the system dynamically adjusts the vehicle speed to 42 km / h, and the braking distance is reduced to 28 meters, ensuring more stable driving; in foggy weather, the system effectively predicts the behavior of pedestrians in front, with a reaction time of 1.2 seconds, and makes obstacle avoidance actions 5 meters in advance to avoid collision risks; the game theory model effectively predicts the deceleration behavior of the vehicle in front, allowing the system to make appropriate deceleration decisions in advance to avoid possible collisions.

[0163] Table 1 Comparison of test data of autonomous driving system

[0164]

[0165] It can be seen from the data in Table 1 above that the method of the present invention performs significantly better than traditional autonomous driving systems in complex environments, can significantly improve the response speed and accuracy of the decision-making process, and can effectively avoid collisions and improve driving safety.

[0166] It can be seen from this embodiment that the autonomous driving system of the present invention can make real-time optimization decisions based on a multi-objective dynamic reward mechanism, a game theory model and a neuroevolutionary algorithm in complex and extreme weather environments. Compared with traditional systems, it has higher decision-making accuracy, stability and safety. Experimental data show that in actual tests, the autonomous driving system of the present invention significantly improves the braking efficiency, obstacle avoidance success rate and reaction speed in rainy, snowy and foggy weather, demonstrating excellent environmental adaptability and reliability.

[0167] The present invention introduces a dynamic reward mechanism based on extreme environment labels. The dynamic reward mechanism can adaptively adjust the reward weight according to different extreme environment types (rainy and snowy weather, heavy fog, night, and complex mountain road conditions). By targetedly strengthening the driving strategy under specific environmental conditions, the present invention can significantly improve the decision-making accuracy and flexibility of the autonomous driving system in complex environments. In rainy and snowy weather, the system sets a higher reward weight for stable driving and braking distance. In foggy weather, the system prioritizes strengthening the obstacle avoidance capability, thereby effectively improving the environmental adaptability and safety of the system. Compared with traditional decision-making methods based on fixed rules or simple machine learning models, the dynamic reward mechanism of the present invention can adjust the decision-making strategy in real time according to environmental changes, ensuring that the autonomous driving system can make safer, more accurate and efficient driving decisions in extreme environments.

[0168] The present invention constructs a multi-factor comprehensive optimization decision model by combining the game theory model to simulate the interaction between vehicles, pedestrians and dynamic obstacles. The game theory model can not only handle complex dynamic interactive behaviors, but also generate multi-objective game strategies based on environmental feature vectors, realizing accurate prediction and optimization of the interactive relationship between multiple actors. In traditional autonomous driving systems, the complexity of these multi-party dynamic interactions is often ignored, resulting in inaccurate or risky decisions in dynamic environments. This problem can be effectively solved by introducing a game theory model, so that vehicles can make more accurate predictions and decisions based on real-time environmental changes when facing multi-party dynamic behaviors, greatly improving the intelligence level and reliability of autonomous driving systems in complex scenarios.

[0169] The present invention adopts a neuroevolutionary game algorithm, and introduces a multi-objective optimized neural network structure to adaptively adjust driving decisions in extreme environments. Different from traditional deep learning methods, the present invention introduces a multi-objective dynamic reward mechanism in the neuroevolution process, and performs real-time weight adjustment in combination with environmental characteristics, thereby optimizing the structure and decision-making ability of the neural network, and being able to dynamically adjust the decision-making strategy according to environmental changes, so that the system can not only quickly adapt and make accurate decisions when dealing with complex and changeable extreme environments, but also effectively avoid the overfitting or underfitting problems of traditional models when the environment changes.

[0170] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A decision-making method for an autonomous driving system in extreme environments based on neuroevolutionary games, characterized in that: The steps include: S1. Collect and construct a real-time environmental data set around the vehicle through the autonomous driving system to obtain weather data, road condition data, and obstacle location data; S2. Input the real-time environmental data set into the data processing unit for data cleaning, noise filtering and feature extraction to obtain environmental feature data, and extract potential obstacle information of different risk levels according to the environmental feature data to generate an environmental feature vector; S3. Based on the environmental feature vector, the environmental feature data is analyzed by the extreme environment recognition module to determine whether the environment in which the vehicle is located belongs to a predefined extreme environment type, which includes rainy and snowy weather, foggy weather, low visibility and complex mountain road conditions; if it is identified as an extreme environment, the extreme environment label is output to the neural evolution module to trigger different multi-objective dynamic reward mechanisms; S4, a multi-objective dynamic reward mechanism corresponding to the activation of extreme environment labels, setting dynamic reward weights for the neural evolution process; S5. Under the guidance of the multi-objective dynamic reward mechanism, the decision optimization model is selected, mutated and reorganized through the neuroevolution model, and the optimal neural network structure and parameter configuration that can adapt to different extreme environments are gradually evolved; S6. Based on the optimal neural network structure, the game theory model is introduced to build a multi-factor comprehensive optimization decision-making model. The game theory model establishes a multi-party behavior prediction model in a complex and extreme environment by simulating the interactive relationship between vehicles, pedestrians and dynamic obstacles, and generates a multi-objective game strategy in combination with the environmental feature vector; S7, using a multi-objective game strategy, which generates the optimal driving path, speed adjustment, obstacle avoidance method and direction control decision based on real-time environmental data and predicted behavior; S8. During the actual driving process, the system inputs the multi-objective game strategy execution feedback data, including vehicle execution results, environmental changes and potential new risk information, into the neural evolution model through real-time feedback data to further adjust the structure and parameter configuration of the neural network.

2. The extreme environment decision-making method of an autonomous driving system based on neuroevolutionary game according to claim 1, characterized in that: The S1 step specifically includes: S11. Using the visual camera in the autonomous driving system to collect real-time image data for detecting visual information in the environment v , using the laser radar in the autonomous driving system to collect point cloud data for generating obstacle information in three-dimensional space l , using millimeter-wave radar in the autonomous driving system to collect target distance and speed information for identifying dynamic objects m , using ultrasonic sensors in the autonomous driving system to collect close-range obstacle data for detecting obstacles at low speed or in a stationary state. u , all sensor data form the initial environment dataset S12. Initial environment data set Perform time series calibration and synchronize the data with the sampling time t to obtain the time-aligned environmental data sequence D env (t); S13, based on environmental data sequence D env (t) data format, decomposing the data into three types of feature information: weather data D weather (t), road condition data D road (t) and obstacle position data D obs (t) and build a real-time environment dataset Among them, f is the multimodal data fusion and feature extraction function.

3. The extreme environment decision-making method of an autonomous driving system based on neuroevolutionary game according to claim 1, characterized in that: The S2 step specifically includes: S21. Real-time environment dataset Input to the data processing unit, perform data cleaning, and generate the cleaned environmental data set D by removing outliers and correcting sampling errors. clean (t); S22, the cleaned environmental data set D clean (t) Perform noise filtering and use a multimodal data joint noise filtering algorithm to process the specific noise characteristics of weather data, road condition data, and obstacle location data, respectively, to generate a noise-filtered environmental data set D filtered (t); S23, environmental data set D after noise filtering filtered (t) Perform feature extraction and extract weather features F through feature extraction algorithm weather (t), road condition characteristics F road (t) and obstacle characteristics F obs (t), generate the environmental feature vector V env (t): V env (t)={F weather (t),F road (t),F obs (t)}。 4. The extreme environment decision-making method of an autonomous driving system based on neuroevolutionary game according to claim 1, characterized in that: The S3 step specifically includes: S31, the environmental feature vector V env (t) Input to the extreme environment recognition module, analyze the weather characteristics, road conditions and obstacle characteristics, and for each predefined extreme environment type E i Calculate each characteristic and extreme environment type E i Matching score: Calculate weather feature matching score in, is the kth weather characteristic index collected at time t, which affects the perception and decision-making ability of the autonomous driving system under extreme weather conditions. In order to map the kth weather characteristic index into a feature mapping function for the matching score of the extreme environment type, the mapping function outputs a high matching score when the extreme environment type is foggy weather and the visibility is low. is the kth weather characteristic index in the extreme environment type E i The importance weight in Calculate the road condition feature matching score in, is the lth road condition characteristic index collected at time t, which affects the driving stability and handling performance of the vehicle. In order to map the lth road condition characteristic index into a characteristic mapping function for the matching score of the extreme environment type, a high matching score is output for the road slope characteristic of the complex mountainous road condition when the slope is steep. is the lth road condition characteristic index in extreme environment type E i The importance weight in Calculate obstacle feature matching score in, is the mth obstacle characteristic index collected at time t, which affects the vehicle’s obstacle avoidance strategy and safety. In order to map the mth obstacle feature index into a feature mapping function for the matching score of the extreme environment type, the obstacle density feature of the low visibility environment outputs a high matching score when the obstacles are dense. is the characteristic index of the mth obstacle in the extreme environment type E i The importance weight in S32. Calculate the extreme environment type E based on the matching scores of each feature i The comprehensive matching score M i (t): Among them, α i , β i , γ i For extreme environment type E i The global weight coefficient under reflects the importance of weather characteristics, road conditions and obstacle characteristics in identifying extreme environment types; S33, the comprehensive matching score M i (t) and the preset extreme environment recognition threshold T i For comparison, if M i (t)≥T i , then the current environment is judged to be an extreme environment type E i , generate the extreme environment label L(t): L(t)={E i ∣M i (t)≥T i }; Where L(t) is the set of extreme environment types identified at time t, including all extreme environment types that meet the conditions, T i It is the identification threshold of extreme environment types, which is set in advance based on experiments and experience.

5. The extreme environment decision-making method of an autonomous driving system based on neuroevolutionary game according to claim 1, characterized in that: The S4 step specifically includes: S41. Activate the corresponding multi-objective dynamic reward mechanism according to the extreme environment label L(t) identified in step S3, and set the dynamic reward weight for the neural evolution process Where i represents the i-th extreme environment type; S42, in rainy and snowy weather conditions E rain ∈L(t), setting the reward weights for stable driving and braking distance For high: Where v(t) is the actual speed of the vehicle at time t, v safe The safe driving speed in rainy and snowy weather is dynamically adjusted according to the slippery degree of the road and the intensity of rainfall. v is the speed deviation tolerance, indicating the permissible range of speed stability, d brake (t) is the actual braking distance of the vehicle at time t, d ideal is the ideal braking distance in rainy and snowy weather, σ d is the tolerance of braking distance deviation, η1 and η2 are reward weight coefficients, which respectively reflect the importance of stable driving and braking distance in rainy and snowy weather; S43, in foggy weather conditions E fog ∈L(t), set the reward weight for obstacle avoidance ability For high: Among them, d obs (t) is the distance between the vehicle and the obstacle in front at time t, d critical is the critical obstacle avoidance distance in foggy weather, λ1 is the distance sensitivity coefficient, which controls the rate at which the obstacle avoidance reward changes with distance, and t response is the obstacle avoidance reaction time of the vehicle at time t, reflecting the agility of the system, t optimal is the optimal obstacle avoidance reaction time in foggy weather, σ t is the reaction time deviation tolerance, which indicates the obstacle avoidance reaction speed requirement, θ1 and θ2 are the reward weight coefficients; S44. In low-visibility nighttime conditions E night ∈L(t), sets the reward weights for object recognition accuracy and reaction speed For high: Among them, N correct (t) is the number of objects correctly identified by the vehicle at time t, N total (t) is the total number of objects actually existing in the environment, Represents the exponential weighting of object recognition accuracy, γ1 is the adjustment index, t decision is the time when the vehicle makes a driving decision at time t, t ideal is the ideal decision response time in nighttime environment, σ τ is the tolerance of decision time deviation, κ1 and κ2 are reward weight coefficients, reflecting the importance of object recognition accuracy and reaction speed in low visibility environment; S45, in complex mountainous road conditions E mountain ∈L(t), giving priority to the reward weights of real-time path planning and obstacle avoidance For high: Among them, C path (t) is the sum of the curvature change rates of the vehicle’s planned path at time t, δ1 is the path complexity sensitivity coefficient, N collision (t) is the number of collisions that occur during the vehicle's obstacle avoidance process at time t, N attempts (t) is the total number of times the vehicle attempts to avoid obstacles, represents the exponential weighting of the obstacle avoidance success rate, β1 is the adjustment index, ξ1 and ξ2 are the reward weight coefficients, which respectively reflect the importance of path planning and obstacle avoidance in complex mountainous road conditions.

6. The extreme environment decision-making method of an autonomous driving system based on neuroevolutionary game according to claim 1, characterized in that: The S6 step specifically includes: S61, based on the optimal neural network structure N(t) and the environmental feature vector V generated in step S5 env (t), combining the extreme environment label L(t) and the state of dynamic objects, a multi-party game environment model is established, where the dynamic objects include vehicles A, pedestrians P, and other dynamic obstacles O; S62, construct the state space S(t), action space A(t) and benefit function U(t) of the game model; S63. Establish a behavior prediction model based on the multi-party game environment model to calculate the optimal strategy π of each game participant A (t),π P (t),π O (t): in, are the optimal strategies for vehicle A, pedestrian P and obstacle O, respectively, P (t and U O (t) is the benefit function of pedestrians and obstacles; S64. Output the optimal strategy generated by the game model to the decision-making module, and generate multi-objective optimized driving decisions in combination with real-time environmental feature data, including speed adjustment, direction control and path selection.

Images