Automatic driving vehicle dynamic environment modeling and active adaptation system

Abstract

The invention relates to automatic driving, in particular to a dynamic environment modeling and active adaptation system for an automatic driving vehicle, which comprises a multi-band cognitive radar module and a millimeter wave radar integrated with multiple bands, adopts an intelligent switching and cooperative working mechanism, and automatically optimizes the duty cycle proportion of the bands according to the environment condition. Radar waves of different frequency bands are transmitted and received, and multi-dimensional information of a target is obtained; the data processing and dynamic environment modeling module adopts a multi-sensor data fusion algorithm to fuse target multi-dimensional information acquired by radars of different frequency bands, performs dynamic environment modeling on the periphery of the vehicle, and updates a dynamic environment model in real time; the decision-making and active adaptation control module is used for making a reasonable driving decision, planning a path and carrying out active adaptation control by adopting a deep learning and artificial intelligence algorithm; according to the technical scheme provided by the invention, the defects that an accurate and real-time dynamic environment model is difficult to construct and the vehicle driving state cannot be timely and reasonably adjusted can be effectively overcome.

Description

Technical Field

[0001] This invention relates to autonomous driving, and more specifically to a dynamic environment modeling and active adaptation system for autonomous vehicles. Background Technology

[0002] In the field of autonomous driving technology, with the rapid development of intelligent transportation systems and the increasing demands for travel safety and convenience, achieving autonomous and safe driving of vehicles in complex environments has become a key goal for the industry. High-precision environmental perception and dynamic environment modeling technologies, as the core support for autonomous driving systems, can provide vehicles with real-time and accurate environmental information, helping them make timely and reasonable decisions. This is of paramount importance for promoting the popularization and application of autonomous driving technology.

[0003] Traditional environmental perception solutions for autonomous vehicles primarily rely on single sensors or simple combinations such as LiDAR, cameras, and single-band millimeter-wave radar. Single-band millimeter-wave radar has limitations in detection range, resolution, and anti-interference capabilities, making it difficult to comprehensively and accurately perceive dynamic targets in various complex environments. While LiDAR can acquire high-precision 3D point cloud data, its high cost limits its large-scale application. Cameras are extremely sensitive to ambient light and weather conditions; in adverse weather (such as rain, fog, and snow) or at night, image quality deteriorates significantly, leading to difficulties in target recognition and failing to reliably provide effective information for autonomous driving.

[0004] Furthermore, existing data fusion methods are relatively simple, failing to fully leverage the advantages of different sensors and exhibiting insufficient data fusion depth, making it difficult to construct accurate and real-time dynamic environment models. In terms of vehicle adaptive control, traditional technologies lack effective analysis and rapid response capabilities for complex environments, failing to adjust vehicle driving states in a timely and reasonable manner, and thus failing to meet the stringent safety and reliability requirements of autonomous driving. These problems severely restrict the development of autonomous driving technology, creating significant safety hazards for vehicles driving in complex traffic scenarios and hindering the transition of autonomous driving technology from the laboratory to practical applications. Summary of the Invention

[0005] (a) Technical problems to be solved In view of the above-mentioned shortcomings of the existing technology, the present invention provides a dynamic environment modeling and active adaptation system for autonomous vehicles, which can effectively overcome the shortcomings of the existing technology, such as the difficulty in constructing accurate and real-time dynamic environment models and the inability to adjust the vehicle driving state in a timely and reasonable manner.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: An autonomous vehicle dynamic environment modeling and active adaptation system includes: The multi-band cognitive radar module integrates millimeter-wave radars of multiple frequency bands. Based on the differences in characteristics of each frequency band radar, it adopts an intelligent switching and collaborative working mechanism, automatically optimizes the frequency band working cycle ratio according to environmental conditions, and transmits and receives radar waves of different frequency bands to obtain multi-dimensional information about the target. The data processing and dynamic environment modeling module receives raw radar data transmitted from the multi-band cognitive radar module, preprocesses the raw radar data, and then uses a multi-sensor data fusion algorithm to fuse multi-dimensional target information acquired by radars of different frequency bands, to model the dynamic environment around the vehicle, and updates the dynamic environment model in real time. The decision-making and active adaptive control module, based on a dynamic environment model, uses deep learning and artificial intelligence algorithms to formulate reasonable driving decisions and plan paths. It actively adapts to control the vehicle by adjusting its speed, direction, and braking, thereby achieving high-precision environmental perception and active adaptive control for autonomous vehicles in dynamic environments.

[0007] Preferably, the multi-band cognitive radar module integrates millimeter-wave radars in multiple frequency bands, including 24GHz, 77GHz, and 79GHz. Based on the differences in characteristics of each frequency band radar: the 24GHz radar utilizes its wide detection range for preliminary detection of long-distance targets, while the 77GHz and 79GHz radars accurately identify the shape and outline of targets due to their high resolution. It adopts an intelligent switching and collaborative working mechanism to automatically optimize the frequency band working cycle ratio according to environmental conditions, and transmits and receives radar waves from different frequency bands to collect multi-dimensional information of the target, including distance, speed, and angle.

[0008] Preferably, the multi-band cognitive radar module incorporates a collaborative detection technology based on dynamic spectrum reconstruction and adaptive beamforming, including: S11. Construct a dynamic spectrum resource allocation model and use reinforcement learning algorithms to optimize the frequency band duty cycle ratio of radars in each frequency band: Define the environment state vector This describes the comprehensive characteristics of the radar's current operating environment, where E represents environmental complexity and T... d For the target density, Interference intensity; Define action space A: ; in, , , The frequency band duty cycles of radars at 24GHz, 77GHz, and 79GHz are respectively. To guide the reinforcement learning algorithm in finding the optimal combination of frequency band duty cycle proportions, a reward function R is designed: ; Where, p det Target detection probability represents the likelihood that the system can detect a target given the current frequency band's operating cycle ratios. res is the comprehensive resolution index, which considers the resolution performance of radars in different frequency bands. tf is the inter-band interference level, measuring the degree of interference between radars in different frequency bands during operation. , , All are weighting coefficients; In the reinforcement learning process, a deep Q-network is used for iterative training, through a neural network. An approximate estimate of the value of performing action A in state S, where For the parameters of the neural network, at each time step t, the algorithm adjusts the parameters based on the state S at the current time step. t Select action A at the current time step. t and perform the action A. t Obtain the state S at the next time step t+1 and the reward R at the current time step t Then by minimizing the loss function Update the parameters of the neural network : ; Among them, A t+1 To determine the state S at the next time step t+1 Choose the action for the next time step. The parameters for the target network are periodically copied from the main network. Update As a discount factor, Represents the mathematical expectation; S12. At the signal processing level, adaptive beamforming technology is introduced to improve the radar's detection performance: For the received signals x of radars in each frequency band i (t), i=24,77,79, the radar receives signals through an array antenna. The array antenna is defined as having N elements. The received signals of the radar in each frequency band are represented as vectors: ; Among them, X i Let x be the received signal vector of the radar in frequency band i. iN (t) represents the received signal of the Nth element of the radar in frequency band i, where t is time; To enhance the target signal and suppress interference signals, the adaptive weighting vector w is calculated using the Minimum Variance Distortionless Response (MVDR) criterion. i First, estimate the covariance matrix of the received signal. ,in Represents the mathematical expectation. For X i The conjugate transpose of the target guidance vector a i This describes the phase relationship of each array element when the target signal arrives at the array antenna, and calculates the adaptive weighting vector according to the MVDR criterion. : ; In practical applications, due to the covariance matrix R i Since it is unknown, an estimate is obtained using finite-time sample data. : , Let M be the received signal vector of the radar at frequency band i at the m-th sample time, where M is the number of samples; By dynamically adjusting the adaptive weighting vector w i This ensures that the main lobe of the radar beam is always aligned with the target signal direction, while simultaneously creating nulls in the direction of interference signals, effectively suppressing sidelobe interference and achieving coordinated enhancement of multi-band radar signals. S13, Received signals x from radars in each frequency band i (t) is weighted and fused to obtain the fused received signal x. fusion (t): .

[0009] Preferably, the data processing and dynamic environment modeling module includes a signal preprocessing unit and an environment modeling unit; The signal preprocessing unit receives the raw radar data transmitted by the multi-band cognitive radar module and uses preprocessing methods including denoising, filtering and calibration to remove interference signals and noise, correct data deviations and improve data accuracy and reliability. The environment modeling unit employs a multi-sensor data fusion algorithm to deeply fuse the data transmitted by the signal preprocessing unit. It combines multi-dimensional information, including the distance, speed, and angle of the target obtained from radars of different frequency bands, to perform dynamic environment modeling around the vehicle and update the dynamic environment model in real time. This enables accurate tracking of dynamic targets and prediction of their motion trajectories, providing real-time and accurate environmental information for autonomous driving.

[0010] Preferably, the environment modeling unit incorporates an improved spatiotemporal adaptive weighted multi-sensor data fusion algorithm, including: Let the measurement value of the target parameter by the i-th sensor among n sensors at time t be x. i(t), i=1,2,…,n, and the covariance matrix of the measurement error is P. i (t); First, define the spatiotemporal correlation coefficient matrix C(t), whose elements c ij (t) represents the spatiotemporal correlation between the measurements of the target parameter by the i-th sensor and the j-th sensor at time t, which is determined by calculating the Euclidean distance between the measurement data of the two sensors in time and space: ; in, , , respectively, are the measurement values ​​of the i-th sensor and the j-th sensor at time t for the k-th target parameter, where m is the dimension of the target parameter; Then, the adaptive weights of the i-th sensor are calculated. : ; Finally, the estimated values ​​of the fused target parameters for: .

[0011] Preferably, a deep learning model based on a hybrid structure of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) is constructed. The pre-processed multi-dimensional target information acquired by radar in different frequency bands is converted into image or sequence format data as model input to further improve the accuracy of data fusion. For image format data, the powerful feature extraction capabilities of Convolutional Neural Networks (CNNs) are used to extract spatial features, including the shape and contour of the target. For sequence-formatted data, the time-series information is processed through a Long Short-Term Memory (LSTM) network to learn the time dependencies between target motions. The model output is an estimate of the fused target parameters. The model is trained using a large amount of real-world scene data to optimize network parameters, enabling the model to accurately fuse multi-band radar data.

[0012] Preferably, the environment modeling unit performs dynamic environment modeling around the vehicle and updates the dynamic environment model in real time, including: S21. Dynamic environment modeling around the vehicle based on an improved octree: 1) Improve the octree to an adaptive partitioning algorithm based on information entropy. Let the i-th octree node O i The covered spatial region contains N i Target data points Each target data point contains information including the target's type, location, and velocity, and an octree node O is defined. i Information entropy for: ; Where, p ik For octree node O i The probability of the k-th target type. n ik For octree node O i The number of data points of the kth target type, where K is the number of target types; Set the information entropy threshold H th When octree node O i Information entropy H(O) i () greater than the information entropy threshold H th At that time, for octree node O i Perform octree partitioning; otherwise, stop partitioning. This adaptive partitioning method allows the octree structure to better fit the environmental characteristics and improves modeling efficiency. 2) Considering the continuity of target motion and the uncertainty of sensor measurements, the calculation method for the occupancy probability of occupancy nodes is updated and optimized by introducing a time decay factor. , and measurement confidence factor octree node O at time t i Occupancy probability p(O) i ,t) is calculated using the following formula: ; Wherein, p(O i (t-1) represents the octree node O at time t-1. i The occupancy probability, n ij Let time t be the time at node O in the octree. i The number of targets detected by the j-th sensor, and the measurement confidence factor. This indicates that time t is at node O of the octree. i The confidence level of the target being detected by the j-th sensor is given, where M is the number of sensors involved in the detection. This formula for calculating the occupancy probability of an occupancy node retains historical occupancy information while dynamically adjusting it based on the current measurement, making the occupancy probability more accurately reflect the environmental state. S22. Real-time updating of dynamic environment model based on dynamic Bayesian network: 1) To more accurately describe the dependencies between octree nodes, a hierarchical dynamic Bayesian network (H-DBN) is introduced, which divides octree nodes into different subnets according to the hierarchy. There are horizontal dependencies between octree nodes in the same subnet and vertical dependencies between octree nodes in different subnets. Let octree node O iLocated in the l-th layer subnet, its state transition probability depends not only on its own state at the previous time step and the states of neighboring nodes in the same layer subnet, but also on the state of the corresponding parent node in the next higher layer subnet. For a three-layer dynamic Bayesian network H-DBN, the octree node O i The state transition probability is: ; in, For the position located at the octree nodes in layered subnet , For the position located at the Regarding octree nodes within the layered subnet The set of adjacent nodes, For the position located at the -1 layer subnet and octree nodes The corresponding set of parent nodes, where t represents the current time and t-1 represents the previous time, this hierarchical structure can process large-scale node data more efficiently and improve the efficiency of dynamic environment model updates; for The calculation constructs a measurement probability model based on the Gaussian Mixture Model (GMM), specifying the radar measurement data z. t In the octree node The state is a mixture of L Gaussian distributions: ; in, This indicates that at time t, given the octree node... Radar measurement data z under the condition t The probability of measurement, Represents radar measurement data z t In the octree node Under the condition of following the mean The covariance matrix is Gaussian distribution, Represents radar measurement data z t In octree node O i The weights, which follow the l-th Gaussian distribution under each state, are estimated by training with Expectation-Maximization (EM) on historical measurement data. , and This allows for more accurate calculation of measurement probabilities and improves the accuracy of dynamic environment model updates.

[0013] Preferably, the decision-making and proactive adaptive control module includes a decision-making planning unit and a control execution unit; The decision-making and planning unit, based on a dynamic environment model, uses deep learning and artificial intelligence algorithms to comprehensively analyze the current environmental conditions, vehicle status, preset safety strategies and driving goals, formulate reasonable driving decisions and plan paths, and quickly generate response strategies when obstacles or dangerous situations are detected ahead, providing decision guidance for vehicle driving. The control execution unit receives driving decisions and path planning instructions transmitted by the decision planning unit, controls the vehicle's power, steering, and braking systems, and performs active adaptive control by adjusting the vehicle's speed, direction, and braking to achieve high-precision environmental perception and active adaptive control of autonomous vehicles in dynamic environments.

[0014] Preferably, the decision planning unit introduces a decision algorithm based on the fusion of graph neural networks (GNN) and reinforcement learning, including: A dynamic graph structure G=(V,E) is constructed based on octree nodes, where the node set V corresponds to the octree nodes and the edge set E represents the adjacency relationship between the octree nodes. For octree node O i Define the node feature vector x i This includes information such as the node's occupancy probability, target type, and location. Graph Neural Networks (GNNs) update node features by aggregating information from neighboring nodes. ; in, octree node In the +1 layer hidden state, For neighboring octree node O j The hidden state at layer l , These are the weight matrix and bias vector of the learnable l-th layer, respectively, and N(i) is the octree node O. i The set of neighboring nodes, For activation functions; Using the node features output by the graph neural network (GNN) as the state input for reinforcement learning, we define the state space S as the set of all node feature vectors, the action space A as the set of driving decisions, and the reward function R(s,a) as follows: ; Where R(s,a) is the reward obtained by performing action a in state s, s' is the next state after performing action a, Safety(s') is a parameter that measures the safety level of the vehicle in state s', calculated based on the distance to surrounding obstacles, Cost(a) is the cost parameter for performing action a, and Progress(s') is a parameter that measures the progress of the vehicle towards the driving target in state s'. , , All are weighting coefficients; Through policy gradient training, the agent learns the optimal driving decision. : ; Wherein, R(s) t ,a t ) represents the state s at the current time step. t Execute action a at the current time step t The rewards received Here, t is the discount factor, and t is the time step. Represents the mathematical expectation. This is a general driving decision.

[0015] Preferably, the control execution unit incorporates a model predictive control and game theory-based collaborative control algorithm, including: Establish a vehicle motion model, and define the vehicle's state vector at time t as... , where x t and y t For position coordinates, v is the driving angle. t For velocity, the control input vector of the vehicle at time t is: , where o t For acceleration, Let be the steering angle. The vehicle's dynamic model is represented as: ; in, Let f be the state of the vehicle at time t+1, and f be the kinematic equation. This is process noise; Construct an optimization problem with prediction in the time domain N, and the objective function J is: ; in, To predict the state of the vehicle at time t+k at time t, The desired state of the vehicle at time t+k To predict the vehicle's control input vector at time t+k-1 at time t, and All are weight vectors; Considering the behavior of surrounding dynamic targets, we introduce the Nash equilibrium from game theory. Suppose there are M dynamic targets around the vehicle, and the strategy of target j is... For j=1,2,…,M, the vehicle's strategy is... Define utility function Solve for the Nash equilibrium: ; in, , ,…, For Nash equilibrium solutions, This is the optimal strategy for the vehicle. ,…, The optimal strategy for dynamic objective j is... This represents the sum of the utility of the vehicle and all participants in its M surrounding dynamic targets. For the strategy of the i-th participant, The strategy combination for all participants except the i-th participant; By using the Nash equilibrium solution as a constraint for model predictive control, the control input of the vehicle is optimized, achieving safer and more efficient active adaptive control.

[0016] (III) Beneficial Effects Compared with the prior art, the dynamic environment modeling and active adaptation system for autonomous vehicles provided by the present invention has the following beneficial effects: 1) At the environmental perception level, the multi-band cognitive radar module integrates millimeter-wave radars in multiple frequency bands such as 24GHz, 77GHz and 79GHz. With the help of intelligent switching and collaborative working mechanism, it gives full play to the advantages of each frequency band radar, makes up for the deficiencies of single frequency band radar in terms of detection distance, resolution and anti-interference capability, and accurately and comprehensively collects multi-dimensional information of the target, including distance, speed and angle. 2) The data processing and dynamic environment modeling module uses data preprocessing and multi-sensor data fusion algorithms to build and update the dynamic environment model in real time, so as to achieve accurate tracking of dynamic targets and prediction of motion trajectories, and provide real-time and accurate environmental information for autonomous driving. 3) The decision-making and active adaptive control module is based on a dynamic environment model and uses deep learning and artificial intelligence algorithms to formulate reasonable driving decisions and plan paths. It actively adapts to control the vehicle by adjusting its speed, direction and braking, thereby achieving high-precision environmental perception and active adaptive control of autonomous vehicles in dynamic environments. This invention significantly improves the accuracy of environmental perception, adaptive control capability, and driving safety of autonomous vehicles in complex environments, effectively reduces accident risks, provides a reliable solution for the development and application of autonomous driving technology, and powerfully promotes the autonomous driving industry to a higher level. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Figure 1 This is a schematic diagram of the first system of the present invention; Figure 2 This is a schematic diagram of the second system of the present invention; Figure 3 This is a schematic diagram of the collaborative detection technology based on dynamic spectrum reconstruction and adaptive beamforming in this invention. Figure 4 This is a schematic diagram of the process of constructing and updating the dynamic environment model in this invention; Figure 5 This is a schematic diagram of the decision-making algorithm based on the fusion of graph neural network (GNN) and reinforcement learning in this invention. Figure 6 This is a schematic diagram of the process for active adaptive control of a vehicle in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] The following describes the specific functional modules of the autonomous vehicle dynamic environment modeling and active adaptation system provided by this invention, using concrete examples (such as...). Figure 1 As shown in the figure, and considering the technical effects, the system functional modules include: The multi-band cognitive radar module integrates millimeter-wave radars of multiple frequency bands. Based on the differences in characteristics of each frequency band radar, it adopts an intelligent switching and collaborative working mechanism, automatically optimizes the frequency band working cycle ratio according to environmental conditions, and transmits and receives radar waves of different frequency bands to obtain multi-dimensional information about the target. The data processing and dynamic environment modeling module receives raw radar data transmitted from the multi-band cognitive radar module, preprocesses the raw radar data, and then uses a multi-sensor data fusion algorithm to fuse multi-dimensional target information acquired by radars of different frequency bands, to model the dynamic environment around the vehicle, and updates the dynamic environment model in real time. The decision-making and active adaptive control module, based on a dynamic environment model, uses deep learning and artificial intelligence algorithms to formulate reasonable driving decisions and plan paths. It actively adapts to control the vehicle by adjusting its speed, direction, and braking, thereby achieving high-precision environmental perception and active adaptive control for autonomous vehicles in dynamic environments.

[0021] I. Multi-band cognitive radar module (i.e. Figure 2 (Data acquisition unit in the middle) The multi-band cognitive radar module integrates millimeter-wave radars in multiple frequency bands, including 24GHz, 77GHz, and 79GHz. Based on the differences in characteristics of each frequency band radar: the 24GHz radar uses its wide detection range to perform preliminary detection of long-distance targets, while the 77GHz and 79GHz radars accurately identify the shape and outline of targets due to their high resolution. It adopts an intelligent switching and collaborative working mechanism to automatically optimize the frequency band working cycle ratio according to environmental conditions, and transmits and receives radar waves of different frequency bands to collect multi-dimensional information of the target, including distance, speed, and angle.

[0022] like Figure 3 As shown, the multi-band cognitive radar module introduces a collaborative detection technology based on dynamic spectrum reconstruction and adaptive beamforming, including: S11. Construct a dynamic spectrum resource allocation model and use reinforcement learning algorithms to optimize the frequency band duty cycle ratio of radars in each frequency band: Define the environment state vector This describes the comprehensive characteristics of the radar's current operating environment, where E represents environmental complexity and T... d For the target density, Interference intensity; Define action space A: ; in, , , The frequency band duty cycles of radars at 24GHz, 77GHz, and 79GHz are respectively. To guide the reinforcement learning algorithm in finding the optimal combination of frequency band duty cycle proportions, a reward function R is designed: ; Where, p det Target detection probability represents the likelihood that the system can detect a target given the current frequency band's operating cycle ratios. res is the comprehensive resolution index, which considers the resolution performance of radars in different frequency bands. tf is the inter-band interference level, measuring the degree of interference between radars in different frequency bands during operation. , , All are weighting coefficients; In the reinforcement learning process, a deep Q-network is used for iterative training, through a neural network. An approximate estimate of the value of performing action A in state S, where For the parameters of the neural network, at each time step t, the algorithm adjusts the parameters based on the state S at the current time step. t Select action A at the current time step. t and perform the action A. t Obtain the state S at the next time step t+1 and the reward R at the current time step t Then by minimizing the loss function Update the parameters of the neural network : ; Among them, A t+1 To determine the state S at the next time step t+1 Choose the action for the next time step. The parameters for the target network are periodically copied from the main network. Update As a discount factor, Represents the mathematical expectation; S12. At the signal processing level, adaptive beamforming technology is introduced to improve the radar's detection performance: For the received signals x of radars in each frequency band i (t), i=24,77,79, the radar receives signals through an array antenna. The array antenna is defined as having N elements. The received signals of the radar in each frequency band are represented as vectors: ; Among them, X i Let x be the received signal vector of the radar in frequency band i. iN (t) represents the received signal of the Nth element of the radar in frequency band i, where t is time; To enhance the target signal and suppress interference signals, the adaptive weighting vector w is calculated using the Minimum Variance Distortionless Response (MVDR) criterion. i First, estimate the covariance matrix of the received signal. ,in Represents the mathematical expectation. For X i The conjugate transpose of the target guidance vector a i This describes the phase relationship of each array element when the target signal arrives at the array antenna, and calculates the adaptive weighting vector according to the MVDR criterion. : ; In practical applications, due to the covariance matrix Ri Since it is unknown, an estimate is obtained using finite-time sample data. : , Let M be the received signal vector of the radar at frequency band i at the m-th sample time, where M is the number of samples; By dynamically adjusting the adaptive weighting vector w i This ensures that the main lobe of the radar beam is always aligned with the target signal direction, while simultaneously creating nulls in the direction of interference signals, effectively suppressing sidelobe interference and achieving coordinated enhancement of multi-band radar signals. S13, Received signals x from radars in each frequency band i (t) is weighted and fused to obtain the fused received signal x. fusion (t): .

[0023] In terms of environmental perception, the above technical solution integrates millimeter-wave radars in multiple frequency bands such as 24GHz, 77GHz, and 79GHz. With intelligent switching and collaborative working mechanisms, it fully leverages the advantages of each frequency band radar, making up for the deficiencies of single-frequency radar in terms of detection distance, resolution, and anti-interference capabilities, and accurately and comprehensively collects multi-dimensional information of the target, including distance, speed, and angle.

[0024] II. Data Processing and Dynamic Environment Modeling Module like Figure 2 As shown, the data processing and dynamic environment modeling module includes a signal preprocessing unit and an environment modeling unit; The signal preprocessing unit receives the raw radar data transmitted by the multi-band cognitive radar module and uses preprocessing methods including denoising, filtering and calibration to remove interference signals and noise, correct data deviations and improve data accuracy and reliability. The environment modeling unit employs a multi-sensor data fusion algorithm to deeply fuse the data transmitted by the signal preprocessing unit. It combines multi-dimensional information, including the distance, speed, and angle of the target obtained from radars of different frequency bands, to model the dynamic environment around the vehicle and update the dynamic environment model in real time. This enables accurate tracking and trajectory prediction of dynamic targets (such as other vehicles, pedestrians, and obstacles), providing real-time and accurate environmental information for autonomous driving.

[0025] like Figure 4 As shown, the environmental modeling unit introduces an improved spatiotemporal adaptive weighted multi-sensor data fusion algorithm, including: Let the measurement value of the target parameter by the i-th sensor among n sensors at time t be x. i (t), i=1,2,…,n, and the covariance matrix of the measurement error is P. i(t); First, define the spatiotemporal correlation coefficient matrix C(t), whose elements c ij (t) represents the spatiotemporal correlation between the measurements of the target parameter by the i-th sensor and the j-th sensor at time t, which is determined by calculating the Euclidean distance between the measurement data of the two sensors in time and space: ; in, , , respectively, are the measurement values ​​of the i-th sensor and the j-th sensor at time t for the k-th target parameter, where m is the dimension of the target parameter; Then, the adaptive weights of the i-th sensor are calculated. : ; Finally, the estimated values ​​of the fused target parameters for: .

[0026] In the technical solution of this application, a deep learning model based on a hybrid structure of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) can be constructed. This model converts pre-processed multi-dimensional target information acquired by radar in different frequency bands into image or sequence format data as input to the model, thereby further improving the accuracy of data fusion. For image format data, the powerful feature extraction capabilities of Convolutional Neural Networks (CNNs) are used to extract spatial features, including the shape and contour of the target. For sequence-formatted data, the time-series information is processed through a Long Short-Term Memory (LSTM) network to learn the time dependencies between target motions. The model output is an estimate of the fused target parameters. The model is trained using a large amount of real-world scene data to optimize network parameters, enabling the model to accurately fuse multi-band radar data.

[0027] like Figure 4 As shown, the environment modeling unit performs dynamic environment modeling around the vehicle and updates the dynamic environment model in real time, including: S21. Dynamic environment modeling around a vehicle based on an improved octree: 1) The improved octree is an adaptive partitioning algorithm based on information entropy. Let the i-th octree node O i The covered spatial region contains N i Target data points Each target data point contains information including the target's type, location, and velocity, and an octree node O is defined. i Information entropy for: ; Where, p ik For octree node O i The probability of the k-th target type. n ik For octree node O i The number of data points of the kth target type, where K is the number of target types; Set the information entropy threshold H th When octree node O i Information entropy H(O) i () greater than the information entropy threshold H th At that time, for octree node O i Perform octree partitioning, or stop partitioning. This adaptive partitioning method allows the octree structure to better fit the environmental characteristics and improves modeling efficiency. 2) Considering the continuity of target motion and the uncertainty of sensor measurements, the calculation method for the occupancy probability of occupancy nodes is updated and optimized by introducing a time decay factor. , and measurement confidence factor octree node O at time t i Occupancy probability p(O) i ,t) is calculated using the following formula: ; Wherein, p(O i (t-1) represents the octree node at time t-1. The occupancy probability, n ij Let time t be the time at octree node O i The number of targets detected by the j-th sensor, and the measurement confidence factor. This indicates that time t is at node t in the octree. The confidence level of the target being detected by the j-th sensor is given, where M is the number of sensors involved in the detection. This formula for calculating the occupancy probability of an occupancy node retains historical occupancy information while dynamically adjusting it based on the current measurement, making the occupancy probability more accurately reflect the environmental state. S22. Real-time updating of dynamic environment model based on dynamic Bayesian network: 1) To more accurately describe the dependencies between octree nodes, a hierarchical dynamic Bayesian network (H-DBN) is introduced, which divides octree nodes into different subnets according to the hierarchy. There are horizontal dependencies between octree nodes in the same subnet and vertical dependencies between octree nodes in different subnets. Let octree node O iLocated in the l-th layer subnet, its state transition probability depends not only on its own state at the previous time step and the states of neighboring nodes in the same layer subnet, but also on the state of the corresponding parent node in the next higher layer subnet. For a three-layer dynamic Bayesian network H-DBN, the octree node O i The state transition probability is: ; in, For the position located at the octree nodes in layered subnet , For the position located at the Regarding octree nodes within the layered subnet The set of adjacent nodes, For the position located at the -1 layer subnet and octree nodes The corresponding set of parent nodes, where t represents the current time and t-1 represents the previous time, this hierarchical structure can process large-scale node data more efficiently and improve the efficiency of dynamic environment model updates; for The calculation constructs a measurement probability model based on the Gaussian Mixture Model (GMM), specifying the radar measurement data z. t In the octree node The state is a mixture of L Gaussian distributions: ; in, This indicates that at time t, given the octree node... Radar measurement data z under the condition t The probability of measurement, Represents radar measurement data z t In octree node O i Under the condition of following the mean The covariance matrix is Gaussian distribution, Represents radar measurement data z t In the octree node Under the state of obeying the first The weights are determined by a Gaussian distribution, and the parameters are estimated through Expectation-Maximization (EM) training on historical measurement data. , and This allows for more accurate calculation of measurement probabilities and improves the accuracy of dynamic environment model updates.

[0028] The above technical solution employs data preprocessing and multi-sensor data fusion algorithms in the data processing and dynamic environment modeling module to construct and update the dynamic environment model in real time, thereby achieving accurate tracking of dynamic targets and prediction of motion trajectories, and providing real-time and accurate environmental information for autonomous driving.

[0029] III. Decision-making and Active Adaptive Control Module like Figure 2 As shown, the decision-making and proactive adaptive control module includes a decision-making and planning unit and a control execution unit; The decision-making and planning unit, based on a dynamic environment model, uses deep learning and artificial intelligence algorithms to comprehensively analyze the current environmental conditions, vehicle status, preset safety strategies and driving goals, formulate reasonable driving decisions and plan paths, and quickly generate response strategies when obstacles or dangerous situations are detected ahead, providing decision guidance for vehicle driving. The control execution unit receives driving decisions and path planning instructions transmitted by the decision planning unit, controls the vehicle's power, steering, and braking systems, and performs active adaptive control by adjusting the vehicle's speed, direction, and braking to achieve high-precision environmental perception and active adaptive control of autonomous vehicles in dynamic environments.

[0030] like Figure 5 and Figure 6 As shown, the decision planning unit introduces a decision algorithm based on the fusion of graph neural networks (GNN) and reinforcement learning, including: A dynamic graph structure G=(V,E) is constructed based on octree nodes, where the node set V corresponds to the octree nodes and the edge set E represents the adjacency relationship between the octree nodes. For octree node O i Define the node feature vector x i This includes information such as the node's occupancy probability, target type, and location. Graph Neural Networks (GNNs) update node features by aggregating information from neighboring nodes. ; in, octree node In the +1 layer hidden state, Neighboring octree nodes The hidden state at layer l , These are the weight matrix and bias vector of the learnable l-th layer, respectively, and N(i) is the octree node. The set of neighboring nodes, For activation functions; Using the node features output by the graph neural network (GNN) as the state input for reinforcement learning, we define the state space S as the set of all node feature vectors, the action space A as the set of driving decisions, and the reward function R(s,a) as follows: ; Where R(s,a) is the reward obtained by performing action a in state s, s' is the next state after performing action a, Safety(s') is a parameter that measures the safety level of the vehicle in state s', calculated based on the distance to surrounding obstacles, Cost(a) is the cost parameter for performing action a, and Progress(s') is a parameter that measures the progress of the vehicle towards the driving target in state s'. , , All are weighting coefficients; Through policy gradient training, the agent learns the optimal driving decision. : ; Wherein, R(s) t ,a t ) represents the state s at the current time step. t Execute action a at the current time step t The rewards received Here, t is the discount factor, and t is the time step. Represents the mathematical expectation. This is a general driving decision.

[0031] like Figure 6 As shown, the control execution unit introduces a collaborative control algorithm based on model predictive control and game theory, including: Establish a vehicle motion model, and define the vehicle's state vector at time t as... , where x t and y t For position coordinates, v is the driving angle. t For velocity, the control input vector of the vehicle at time t is: , where o t For acceleration, Let be the steering angle. The vehicle's dynamic model is represented as: ; in, Let f be the state of the vehicle at time t+1, and f be the kinematic equation. This is process noise; Construct an optimization problem with prediction in the time domain N, and the objective function J is: ; in, To predict the state of the vehicle at time t+k at time t, Let be the desired state of the vehicle at time t+k. To predict the vehicle's control input vector at time t+k-1 at time t, and Both are weight vectors; Considering the behavior of surrounding dynamic targets, we introduce the Nash equilibrium from game theory. Suppose there are M dynamic targets around the vehicle, and the strategy of target j is... For j=1,2,…,M, the vehicle's strategy is... Define utility function Solve for the Nash equilibrium: ; in, , ,…, For Nash equilibrium solutions, This is the optimal strategy for the vehicle. ,…, The optimal strategy for dynamic objective j is... This represents the sum of the utility of the vehicle and all participants in its M surrounding dynamic targets. For the strategy of the i-th participant, The strategy combination for all participants except the i-th participant;

[0032] By using the Nash equilibrium solution as a constraint for model predictive control, the control input of the vehicle is optimized, achieving safer and more efficient active adaptive control.

[0033] The above technical solution uses a decision-making and active adaptive control module based on a dynamic environment model. It employs deep learning and artificial intelligence algorithms to formulate reasonable driving decisions and plan paths. By adjusting the vehicle's speed, direction, and braking, it performs active adaptive control, thereby achieving high-precision environmental perception and active adaptive control for autonomous vehicles in dynamic environments.

[0034] Based on the aforementioned disclosure of an autonomous vehicle dynamic environment modeling and active adaptation system, this invention also provides an autonomous vehicle dynamic environment modeling and active adaptation device, including a memory, a processor, and a computer program stored in the memory and running in the processor. When the processor executes the computer program, it employs the method corresponding to the aforementioned autonomous vehicle dynamic environment modeling and active adaptation system to achieve high-precision environmental perception and active adaptive control of the autonomous vehicle.

[0035] The autonomous vehicle dynamic environment modeling and active adaptation device provided in this embodiment is essentially a computer device used to implement the above-mentioned system and its corresponding method. This computer device can be an embedded device installed in each vehicle. This embedded device can communicate and interact with other devices through a server to obtain the required data and perform data processing using independent computing or edge computing methods. Alternatively, this computer device can be a backend device independent of the vehicle, completing the data processing process when a user makes a request, and sending dynamic environment modeling information to the data terminals in each vehicle based on the data processing results, thus assisting the vehicle in completing active adaptive control.

[0036] The computer equipment in this embodiment can also be a smart terminal, tablet computer, laptop computer, desktop computer, rack server, blade server, tower server or cabinet server (including independent servers or server clusters composed of multiple servers), etc.

[0037] The computer device in this embodiment includes, but is not limited to, a memory and a processor that can be interconnected via a system bus. The memory (i.e., readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (such as SD, DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as the hard disk or RAM of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Of course, the memory can also include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. Furthermore, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0038] In some embodiments, a processor may be a central processing unit (CPU), a graphics processing unit (GPU), a controller, a microcontroller, a microprocessor, a data processing chip, etc. This processor is typically used to control the overall operation of a computer device. In this embodiment, the processor is used to run program code stored in memory or process data.

[0039] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic environment modeling and active adaptation system for autonomous vehicles, characterized in that: include: The multi-band cognitive radar module integrates millimeter-wave radars of multiple frequency bands. Based on the differences in characteristics of each frequency band radar, it adopts an intelligent switching and collaborative working mechanism, automatically optimizes the frequency band working cycle ratio according to environmental conditions, and transmits and receives radar waves of different frequency bands to obtain multi-dimensional information about the target. The data processing and dynamic environment modeling module receives raw radar data transmitted from the multi-band cognitive radar module, preprocesses the raw radar data, and then uses a multi-sensor data fusion algorithm to fuse multi-dimensional target information acquired by radars of different frequency bands, to model the dynamic environment around the vehicle, and updates the dynamic environment model in real time. The decision-making and active adaptive control module, based on a dynamic environment model, uses deep learning and artificial intelligence algorithms to formulate reasonable driving decisions and plan paths. It actively adapts to control the vehicle by adjusting its speed, direction, and braking, thereby achieving high-precision environmental perception and active adaptive control for autonomous vehicles in dynamic environments.

2. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 1, characterized in that: The multi-band cognitive radar module integrates millimeter-wave radars in multiple frequency bands, including 24GHz, 77GHz, and 79GHz. Based on the differences in characteristics of each frequency band radar: the 24GHz radar utilizes its wide detection range for preliminary detection of long-distance targets, while the 77GHz and 79GHz radars accurately identify the shape and outline of targets due to their high resolution. It adopts an intelligent switching and collaborative working mechanism to automatically optimize the frequency band working cycle ratio according to environmental conditions, and transmits and receives radar waves from different frequency bands to collect multi-dimensional information of the target, including distance, speed, and angle.

3. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 2, characterized in that: The multi-band cognitive radar module introduces a collaborative detection technology based on dynamic spectrum reconstruction and adaptive beamforming, including: S11. Construct a dynamic spectrum resource allocation model and use reinforcement learning algorithms to optimize the frequency band duty cycle ratio of radars in each frequency band: Define the environment state vector This describes the comprehensive characteristics of the radar's current operating environment, where E represents environmental complexity and T... d For the target density, Interference intensity; Define action space A: ； in, , , The frequency band duty cycles of radars at 24GHz, 77GHz, and 79GHz are respectively. To guide the reinforcement learning algorithm in finding the optimal combination of frequency band duty cycle proportions, a reward function R is designed: ； Where, p det Target detection probability represents the likelihood that the system can detect a target given the current frequency band's operating cycle ratios. res is the comprehensive resolution index, which considers the resolution performance of radars in different frequency bands. tf is the inter-band interference level, measuring the degree of interference between radars in different frequency bands during operation. , , All are weighting coefficients; In the reinforcement learning process, a deep Q-network is used for iterative training, through a neural network. An approximate estimate of the value of performing action A in state S, where For the parameters of the neural network, at each time step t, the algorithm adjusts the parameters based on the state S at the current time step. t Select action A at the current time step. t and perform the action A. t Obtain the state S at the next time step t+1 and the reward R at the current time step t Then by minimizing the loss function Update the parameters of the neural network : ； Among them, A t+1 To determine the state S at the next time step t+1 Choose the action for the next time step. The parameters for the target network are periodically copied from the main network. Update As a discount factor, Represents the mathematical expectation; S12. At the signal processing level, adaptive beamforming technology is introduced to improve the radar's detection performance: For the received signals x of radars in each frequency band i (t), i=24,77,79, the radar receives signals through an array antenna. The array antenna is defined as having N elements. The received signals of the radar in each frequency band are represented as vectors: ； Among them, X i Let x be the received signal vector of the radar in frequency band i. iN (t) represents the received signal of the Nth element of the radar in frequency band i, where t is time; To enhance the target signal and suppress interference signals, the adaptive weighting vector w is calculated using the Minimum Variance Distortionless Response (MVDR) criterion. i First, estimate the covariance matrix of the received signal. ,in Represents the mathematical expectation. For X i The conjugate transpose of the target guidance vector a i This describes the phase relationship of each array element when the target signal arrives at the array antenna, and calculates the adaptive weighting vector according to the MVDR criterion. : ； In practical applications, due to the covariance matrix R i Since it is unknown, an estimate is obtained using finite-time sample data. : , Let M be the received signal vector of the radar at frequency band i at the m-th sample time, where M is the number of samples; By dynamically adjusting the adaptive weighting vector w i This ensures that the main lobe of the radar beam is always aligned with the target signal direction, while simultaneously creating nulls in the direction of interference signals, effectively suppressing sidelobe interference and achieving coordinated enhancement of multi-band radar signals. S13, Received signals x from radars in each frequency band i (t) is weighted and fused to obtain the fused received signal x. fusion (t): 。 4. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 1, characterized in that: The data processing and dynamic environment modeling module includes a signal preprocessing unit and an environment modeling unit. The signal preprocessing unit receives the raw radar data transmitted by the multi-band cognitive radar module and uses preprocessing methods including denoising, filtering and calibration to remove interference signals and noise, correct data deviations and improve data accuracy and reliability. The environment modeling unit employs a multi-sensor data fusion algorithm to deeply fuse the data transmitted by the signal preprocessing unit. It combines multi-dimensional information, including the distance, speed, and angle of the target obtained from radars of different frequency bands, to perform dynamic environment modeling around the vehicle and update the dynamic environment model in real time. This enables accurate tracking of dynamic targets and prediction of their motion trajectories, providing real-time and accurate environmental information for autonomous driving.

5. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 4, characterized in that: The environmental modeling unit introduces an improved spatiotemporal adaptive weighted multi-sensor data fusion algorithm, including: Let the measurement value of the target parameter by the i-th sensor among n sensors at time t be x. i (t), i=1,2,…,n, and the covariance matrix of the measurement error is P. i (t); First, define the spatiotemporal correlation coefficient matrix C(t), whose elements c ij (t) represents the spatiotemporal correlation between the measurements of the target parameter by the i-th sensor and the j-th sensor at time t, which is determined by calculating the Euclidean distance between the measurement data of the two sensors in time and space: ； in, , , respectively, are the measurement values ​​of the i-th sensor and the j-th sensor at time t for the k-th target parameter, where m is the dimension of the target parameter; Then, the adaptive weights of the i-th sensor are calculated. : ； Finally, the estimated values ​​of the fused target parameters for: 。 6. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 5, characterized in that: A deep learning model based on a hybrid structure of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) is constructed. Preprocessed multi-dimensional target information acquired by radar across different frequency bands is transformed into image or sequence data as model input to further improve the accuracy of data fusion. For image format data, the powerful feature extraction capabilities of Convolutional Neural Networks (CNNs) are used to extract spatial features, including the shape and contour of the target. For sequence-formatted data, the time-series information is processed through a Long Short-Term Memory (LSTM) network to learn the time dependencies between target motions. The model output is an estimate of the fused target parameters. The model is trained using a large amount of real-world scene data to optimize network parameters, enabling the model to accurately fuse multi-band radar data.

7. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 5, characterized in that: The environment modeling unit performs dynamic environment modeling around the vehicle and updates the dynamic environment model in real time, including: S21. Dynamic environment modeling around the vehicle based on an improved octree: 1) Improve the octree to an adaptive partitioning algorithm based on information entropy. Let the i-th octree node O i The covered spatial region contains N i Target data points Each target data point contains information including the target's type, location, and velocity, and an octree node O is defined. i Information entropy for: ； Where, p ik For octree node O i The probability of the k-th target type. n ik For octree node O i The number of data points of the kth target type, where K is the number of target types; Set the information entropy threshold H th When octree node O i Information entropy H(O) i () greater than the information entropy threshold H th At that time, for octree node O i Perform octree partitioning, or stop partitioning. This adaptive partitioning method allows the octree structure to better fit the environmental characteristics and improves modeling efficiency. 2) Considering the continuity of target motion and the uncertainty of sensor measurements, the calculation method for the occupancy probability of occupancy nodes is updated and optimized by introducing a time decay factor. , and measurement confidence factor octree node O at time t i Occupancy probability p(O) i ,t) is calculated using the following formula: ； Wherein, p(O i (t-1) represents the octree node O at time t-1. i The occupancy probability, n ij Let time t be the time at octree node O i The number of targets detected by the j-th sensor, and the measurement confidence factor. This indicates that time t is at node O in the octree. i The confidence level of the target being detected by the j-th sensor is given, where M is the number of sensors involved in the detection. This formula for calculating the occupancy probability of an occupancy node retains historical occupancy information while dynamically adjusting it based on the current measurement, making the occupancy probability more accurately reflect the environmental state. S22. Real-time updating of dynamic environment model based on dynamic Bayesian network: 1) To more accurately describe the dependencies between octree nodes, a hierarchical dynamic Bayesian network (H-DBN) is introduced, which divides octree nodes into different subnets according to the hierarchy. There are horizontal dependencies between octree nodes in the same subnet and vertical dependencies between octree nodes in different subnets. Let octree node O i Located in the l-th layer subnet, its state transition probability depends not only on its own state at the previous time step and the states of neighboring nodes in the same layer subnet, but also on the state of the corresponding parent node in the next higher layer subnet. For a three-layer dynamic Bayesian network H-DBN, the octree node O i The state transition probability is: ； in, For the position located at the octree nodes in layered subnet , For the octree node O located in the l-th subnet... i The set of adjacent nodes, For the position located at the -1 layer subnet and octree nodes The corresponding set of parent nodes, where t represents the current time and t-1 represents the previous time, this hierarchical structure can process large-scale node data more efficiently and improve the efficiency of dynamic environment model updates; for The calculation constructs a measurement probability model based on the Gaussian Mixture Model (GMM), specifying the radar measurement data z. t In the octree node The state is a mixture of L Gaussian distributions: ； in, This indicates that at time t, given the octree node... Radar measurement data z under the condition t The probability of measurement, Represents radar measurement data z t In the octree node Under the condition of following the mean The covariance matrix is Gaussian distribution, Represents radar measurement data z t In the octree node Under the state of obeying the first The weights are determined by a Gaussian distribution, and the parameters are estimated through Expectation-Maximization (EM) training on historical measurement data. , and This allows for more accurate calculation of measurement probabilities and improves the accuracy of dynamic environment model updates.

8. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 1, characterized in that: The decision-making and proactive adaptive control module includes a decision-making planning unit and a control execution unit; The decision-making and planning unit, based on a dynamic environment model, uses deep learning and artificial intelligence algorithms to comprehensively analyze the current environmental conditions, vehicle status, preset safety strategies and driving goals, formulate reasonable driving decisions and plan paths, and quickly generate response strategies when obstacles or dangerous situations are detected ahead, providing decision guidance for vehicle driving. The control execution unit receives driving decisions and path planning instructions transmitted by the decision planning unit, controls the vehicle's power, steering, and braking systems, and performs active adaptive control by adjusting the vehicle's speed, direction, and braking to achieve high-precision environmental perception and active adaptive control of autonomous vehicles in dynamic environments.

9. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 8, characterized in that: The decision planning unit introduces a decision algorithm based on the fusion of graph neural networks (GNN) and reinforcement learning, including: A dynamic graph structure G=(V,E) is constructed based on octree nodes, where the node set V corresponds to the octree nodes and the edge set E represents the adjacency relationship between the octree nodes. For octree node O i Define the node feature vector x i This includes information such as the node's occupancy probability, target type, and location. Graph Neural Networks (GNNs) update node features by aggregating information from neighboring nodes. ； in, octree node In the +1 layer hidden state, For neighboring octree node O j The hidden state at layer l , These are the weight matrix and bias vector of the learnable l-th layer, respectively, and N(i) is the octree node O. i The set of neighboring nodes, For activation functions; Using the node features output by the graph neural network (GNN) as the state input for reinforcement learning, we define the state space S as the set of all node feature vectors, the action space A as the set of driving decisions, and the reward function R(s,a) as follows: ； Where R(s,a) is the reward obtained by performing action a in state s, s' is the next state after performing action a, Safety(s') is a parameter that measures the safety level of the vehicle in state s', calculated based on the distance to surrounding obstacles, Cost(a) is the cost parameter for performing action a, and Progress(s') is a parameter that measures the progress of the vehicle towards the driving target in state s'. , , All are weighting coefficients; Through policy gradient training, the agent learns the optimal driving decision. : ； Wherein, R(s) t ,a t ) represents the state s at the current time step. t Execute action a at the current time step t The rewards received Here, t is the discount factor, and t is the time step. Represents the mathematical expectation. This is a general driving decision.

10. The autonomous vehicle dynamic environment modeling and active adaptation system according to claim 9, characterized in that: The control execution unit incorporates a collaborative control algorithm based on model predictive control and game theory, including: Establish a vehicle motion model, and define the vehicle's state vector at time t as... , where x t and y t For position coordinates, v is the driving angle. t For velocity, the control input vector of the vehicle at time t is: , where o t For acceleration, Let be the steering angle. The vehicle's dynamic model is represented as: ； in, Let f be the state of the vehicle at time t+1, and f be the kinematic equation. This is process noise; Construct an optimization problem with prediction in the time domain N, and the objective function J is: ； in, To predict the state of the vehicle at time t+k at time t, Let be the desired state of the vehicle at time t+k. To predict the vehicle's control input vector at time t+k-1 at time t, and Both are weight vectors; Considering the behavior of surrounding dynamic targets, we introduce the Nash equilibrium from game theory. Suppose there are M dynamic targets around the vehicle, and the strategy of target j is... For j=1,2,…,M, the vehicle's strategy is... Define utility function Solve for the Nash equilibrium: ； in, , ,…, For Nash equilibrium solutions, This is the optimal strategy for the vehicle. ,…, The optimal strategy for dynamic objective j is... This represents the sum of the utility of the vehicle and all participants in its M surrounding dynamic targets. For the strategy of the i-th participant, The strategy combination for all participants except the i-th participant; By using the Nash equilibrium solution as a constraint for model predictive control, the control input of the vehicle is optimized, achieving safer and more efficient active adaptive control.

Images