Automatic driving system based on intelligent traffic

Abstract

The invention relates to the technical field of automatic driving, in particular to an automatic driving system based on intelligent traffic, which comprises a multi-source heterogeneous sensing fusion unit, a group collaborative decision-making unit and a dynamic control optimization unit, the group collaborative decision-making unit carries out space-time alignment on multi-source heterogeneous data, constructs and updates a global environment perception knowledge graph, carries out modeling on vehicle trajectories based on a Monte Carlo tree search framework in combination with deep reinforcement learning and risk perception, and trains a deep Q network model by adopting federated learning and adversarial training. A traffic flow distribution scheme is generated by applying multi-agent deep reinforcement learning and a game theory, storage data is proved by using homomorphic encryption and zero knowledge, a dynamic control optimization unit predicts, controls and plans an acceleration curve through a nonlinear model, and a transverse path tracking error is controlled by fusing sliding mode control and a fuzzy logic algorithm. And the safety of automatic driving is improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an autonomous driving system based on intelligent transportation. Background Art

[0002] Autonomous driving is an important technology. In the field of autonomous driving technology, the fusion processing of multi-source data is a key link in achieving safe and efficient autonomous driving, but the traditional multi-source data fusion method has significant defects.

[0003] When traditional methods fuse multi-source heterogeneous data such as environmental point clouds, traffic signal phases, and surrounding vehicle status from different sensors, the data cannot be effectively matched in time and space due to differences in the acquisition frequency, time base, and spatial coordinate system of different sensors. This leads to deviations and errors in the fused data, making it difficult to accurately align them in time and space. In complex intersection scenarios, the traffic signal phase data collected by the camera and the environmental point cloud data obtained by the lidar are not accurately aligned in time and space, making it impossible for the autonomous driving system to accurately judge the vehicle's driving environment and traffic signal status. This may lead to vehicle decision-making errors, safety hazards such as running red lights and colliding with other vehicles, seriously restricting the reliability and safety of the autonomous driving system. To solve this technical problem, we provide an autonomous driving system based on intelligent transportation. Summary of the Invention

[0004] The purpose of the present invention is to provide an automatic driving system based on intelligent transportation to solve the problems raised in the above background technology.

[0005] 1. Traditional multi-source data fusion cannot accurately align multi-source heterogeneous data in time and space, which can lead to decision-making errors in the autonomous driving system and pose safety risks. Therefore, this case uses a multi-source heterogeneous perception fusion unit and an algorithm based on event-driven and multi-scale feature matching to accurately align data, improve the accuracy of the autonomous driving system's environmental judgment, and reduce safety risks.

[0006] 2. Because traditional methods make it difficult to effectively utilize historical data from roadside MEC servers to generate optimal traffic flow allocation plans, this case uses an algorithm based on federated learning and adversarial training through a group collaborative decision-making unit. Each node conducts local training and introduces adversarial training, and the central server aggregates and updates the model. This can fully tap the value of data, optimize traffic flow allocation plans, and improve road traffic efficiency.

[0007] To achieve the above objectives, an autonomous driving system based on intelligent transportation is provided, which includes a multi-source heterogeneous perception fusion unit, a group collaborative decision-making unit, and a dynamic control optimization unit;

[0008] The multi-source heterogeneous perception fusion unit acquires data on environmental point clouds, traffic signal phases, and surrounding vehicle status through a sensor array, and performs spatiotemporal alignment processing on these data to generate a global environmental perception knowledge graph.

[0009] The swarm collaborative decision-making unit performs spatiotemporal joint optimization decisions based on the global environmental perception knowledge graph. It models vehicle trajectories within a 100-meter radius based on a Monte Carlo tree search framework, outputs a collision-free passage sequence, aggregates historical passage data from the roadside MEC server to train a deep Q network model, generates a traffic flow allocation plan, and then stores the collision-free passage sequence and traffic flow allocation plan through a blockchain hash chain.

[0010] The dynamic control optimization unit adopts nonlinear model predictive control to plan acceleration curve based on collision-free passage sequence and traffic flow distribution scheme, and integrates sliding mode control and fuzzy logic algorithm to control lateral path tracking error.

[0011] As a further improvement to this technical solution, the multi-source heterogeneous perception fusion unit uses an adaptive spatiotemporal alignment algorithm based on event-driven and multi-scale feature matching when performing spatiotemporal alignment processing on multimodal data:

[0012] Set specific event trigger conditions for different sensors. When an event is triggered, record the data of each sensor at that moment as keyframe data, process the keyframe data, and extract the features of the environmental point cloud, traffic signal phase, and surrounding vehicle status data respectively;

[0013] A feature matching network is used to match features. Through training, the network learns the mapping relationship between different modal features. Then, the prediction and update steps of the Kalman filter are used, combined with the sensor's measurement error covariance matrix and the result of feature matching, to iteratively correct the temporal and spatial deviations of the environmental point cloud, traffic signal phase, and surrounding vehicle status data.

[0014] As a further improvement of this technical solution, when the multi-source heterogeneous perception fusion unit generates a global environment perception knowledge graph, a knowledge graph construction method based on graph attention network and semantic reasoning is adopted:

[0015] Represent the spatiotemporally aligned data as a graph structure, and assign an initial feature vector to each node and edge in the graph structure. The initial feature vector contains features extracted from the environmental point cloud, traffic signal phase, and surrounding vehicle status data.

[0016] A graph attention network is used to extract and learn the initial feature vectors of the graph structure. By stacking multiple layers of graph attention networks, the features of nodes and edges are fused with information at different levels. Pre-defined semantic rules and ontology knowledge base are then used to perform semantic reasoning on the features that fuse information at different levels. Through semantic reasoning, new relationships and attributes are added to the global environment perception knowledge graph to improve the content of the global environment perception knowledge graph.

[0017] As a further improvement to this technical solution, the multi-source heterogeneous perception fusion unit uses a method based on incremental graph neural network and evidence theory to perform real-time updating and error correction on the generated global environment perception knowledge graph:

[0018] When new data is input, an incremental graph neural network algorithm is used to update the global environment perception knowledge graph. During the update process, an incremental training method is used to adjust the parameters of the incremental graph neural network according to the new data.

[0019] Evidence theory is used to deal with conflicts between different sensor data, and finally a consistency check mechanism is established to perform logical verification on the global environment perception knowledge graph.

[0020] As a further improvement to this technical solution, the group collaborative decision-making unit uses a Monte Carlo tree search algorithm based on deep reinforcement learning and risk perception when modeling vehicle trajectories within a radius of 100 meters based on the Monte Carlo tree search framework:

[0021] The traffic environment within a 100-meter radius is set as a reinforcement learning environment. The state, action, and reward functions of the reinforcement learning environment are defined separately. A risk perception module is introduced at each node of the Monte Carlo tree search. The risk perception module is used to assess the risk between vehicles in real time and input the risk assessment results into the decision-making process.

[0022] During the simulation of Monte Carlo tree search, the agent outputs the action probability distribution through the policy network of the reinforcement learning environment based on the current state and risk assessment results, selects the optimal action, and updates the policy network and value network based on the optimal action.

[0023] As a further improvement to this technical solution, the group collaborative decision-making unit adopts a deep Q network algorithm based on federated learning and adversarial training when aggregating historical traffic data in the roadside MEC server to train the deep Q network model:

[0024] Establish a federated learning network consisting of roadside MEC servers and autonomous vehicles. Each federated learning network node locally stores historical traffic data and trains a deep Q network model locally.

[0025] An adversarial training mechanism is introduced during the local training process of each federated learning network node to optimize the deep Q network model. After local training, each federated learning network node uploads the model parameters to the central server. The central server aggregates the model parameters according to the federated learning aggregation algorithm to obtain the global model parameters, and then sends the global model parameters to each federated learning network node to update the deep Q network model.

[0026] As a further improvement to this technical solution, the group collaborative decision-making unit adopts a traffic flow allocation algorithm based on multi-agent deep reinforcement learning and game theory when generating a traffic flow allocation plan:

[0027] Each autonomous vehicle and roadside infrastructure is considered as a sub-agent, each with its own state and action space;

[0028] Each sub-agent is trained using a deep reinforcement learning algorithm. The sub-agent selects sub-actions based on the current state and updates the sub-agent policy network and sub-agent value network based on feedback from the environment.

[0029] Game theory methods are used in multi-agent systems to optimize the strategies of sub-agent strategy networks and generate the final traffic flow allocation plan.

[0030] As a further improvement to this technical solution, the group collaborative decision-making unit adopts a blockchain storage method based on homomorphic encryption and zero-knowledge proof when storing collision-free passage sequences and traffic flow allocation plans through the blockchain hash chain:

[0031] The collision-free passage sequence and traffic flow allocation plan are homomorphically encrypted, and the encrypted data is verified using zero-knowledge proof technology before being stored in the blockchain. The data that has been homomorphically encrypted and verified by zero-knowledge proof is hashed to generate a hash value, which is then stored in the block of the blockchain.

[0032] As a further improvement of the present technical solution, the method of the dynamic control optimization unit using a nonlinear model to predict and control the acceleration curve based on the collision-free passage sequence and traffic flow distribution scheme is as follows:

[0033] A vehicle dynamics prediction model based on neural network is established, and the vehicle state is predicted according to the vehicle dynamics prediction model. The optimization target is set in the optimization process of model predictive control, and then the optimization algorithm is used to solve the optimization problem of the optimization target to obtain the optimal acceleration curve.

[0034] As a further improvement of the present technical solution, the dynamic control optimization unit integrates sliding mode control and fuzzy logic algorithm to perform lateral path tracking error control, and adopts a lateral path tracking method based on fuzzy sliding mode variable structure control and neural network compensation:

[0035] Design a sliding mode controller that calculates control inputs based on the vehicle's lateral position error and heading error to control the vehicle's lateral motion to achieve the desired path.

[0036] Design a neural network compensator that takes the vehicle's lateral position error, heading error, and control input as input and outputs a compensation for the sliding mode controller.

[0037] The control input and the compensation amount of the sliding mode controller are synthesized to obtain the final control input, which is used to control the vehicle's steering system to perform lateral path tracking.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] In an autonomous driving system based on intelligent transportation, a multi-source heterogeneous perception fusion unit uses a specific algorithm for spatiotemporal alignment, which solves the problem of spatiotemporal matching in traditional data fusion and generates an accurate global environmental perception knowledge graph, enabling the autonomous driving system to more accurately grasp the driving environment and improve driving safety. In the decision-making link, the group collaborative decision-making unit uses a variety of advanced algorithms to incorporate risk perception when modeling vehicle trajectories, uses federated learning to train deep Q network models, and optimizes multi-agent strategies based on game theory to generate traffic flow distribution plans and securely store relevant data. This not only optimizes vehicle traffic sequences and reduces collision risks, but also improves road resource utilization. In terms of vehicle control, the dynamic control optimization unit uses nonlinear model predictive control to plan acceleration curves, and combines sliding mode control with fuzzy logic algorithms to control lateral path tracking errors, which can make the vehicle drive more smoothly and accurately track the path, thereby improving the comfort and stability of autonomous driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a system structure block diagram of the present invention;

[0041] Figure 2 It is the workflow diagram of the present invention.

[0042] The meaning of each number in the figure is:

[0043] 1. Multi-source heterogeneous perception fusion unit; 2. Group collaborative decision-making unit; 3. Dynamic control optimization unit. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] The present invention provides an automatic driving system based on intelligent transportation. Figure 1-Figure 2 As shown, it includes a multi-source heterogeneous perception fusion unit 1, a group collaborative decision-making unit 2, and a dynamic control optimization unit 3;

[0046] The multi-source heterogeneous perception fusion unit 1 obtains environmental point clouds, traffic signal phases, and surrounding vehicle status data through the sensor array, and performs spatiotemporal alignment processing on them to generate a global environmental perception knowledge graph;

[0047] When performing spatiotemporal alignment processing on multimodal data, the multi-source heterogeneous perception fusion unit 1 adopts an adaptive spatiotemporal alignment algorithm based on event-driven and multi-scale feature matching:

[0048] Setting specific event trigger conditions for different sensors can capture the key time points of data changes. Selecting the data at these moments as keyframe data can reduce unnecessary data processing. When an event is triggered, the data of each sensor at that moment is recorded as keyframe data. For lidar, the event is set to be triggered when the speed change of the target object exceeds a certain threshold. For cameras, the event is triggered when the color of the traffic signal is detected. When the event is triggered, the data of each sensor at that moment is recorded as keyframe data, processed, and the features of the environmental point cloud, traffic signal phase, and surrounding vehicle status data are extracted respectively, laying the foundation for more accurate spatiotemporal alignment.

[0049] A feature matching network is constructed. The network structure includes convolutional layers, pooling layers and fully connected layers. The acquired features are used to train the feature matching network, and the features are matched using the feature matching network. The back-propagation algorithm is used to adjust the network weights through training so that the network can learn the mapping relationship between different modal features, establish a state transition model, and predict the position and posture of each modal data at different time steps. The prediction and update steps of the Kalman filter are then used to estimate the state at the current moment based on the state at the previous moment. In the update step, the time and space deviations of the environmental point cloud, traffic signal phase and surrounding vehicle status data are iteratively corrected in combination with the sensor's measurement error covariance matrix and the result of feature matching. The prediction and update steps are repeated and iterative corrections are performed until the time and space deviations meet the preset threshold, providing an accurate data foundation for the subsequent generation of a global environmental perception knowledge graph.

[0050] When the multi-source heterogeneous perception fusion unit 1 generates the global environment perception knowledge graph, it adopts the knowledge graph construction method based on graph attention network and semantic reasoning:

[0051] The spatiotemporally aligned data is represented as a graph structure, with vehicles, traffic lights, and obstacles represented as nodes of the graph and their relationships as edges. An initial feature vector is assigned to each node and edge in the graph structure. The initial feature vector contains features extracted from the environmental point cloud, traffic signal phase, and surrounding vehicle status data.

[0052] A graph attention network is used to extract and learn the initial feature vectors of the graph structure. In the graph attention network, each node calculates the attention weight between itself and adjacent nodes through the attention mechanism, performs weighted summation on the features of adjacent nodes according to the attention weight, and updates its own feature representation. By stacking multiple layers of graph attention networks, the features of nodes and edges are continuously updated, so that the features of nodes and edges are integrated with information at different levels. Pre-defined semantic rules and ontology knowledge base are then used to perform semantic reasoning on the features that integrate information at different levels. Through semantic reasoning, new relationships and attributes are added to the global environment perception knowledge graph, and the content of the global environment perception knowledge graph is improved, so that the knowledge graph has stronger semantic expression capabilities, can more accurately describe environmental information, and provide richer knowledge support for the decision-making of the autonomous driving system.

[0053] When the multi-source heterogeneous perception fusion unit 1 performs real-time updates and error correction on the generated global environment perception knowledge graph, it adopts a method based on incremental graph neural network and evidence theory:

[0054] When new data is input, it determines which nodes and edges are affected and only updates the features of these affected nodes and edges without changing the features of other nodes and edges. During the update process, an incremental training method is used to adjust the parameters of the incremental graph neural network according to the new data.

[0055] A trust level is assigned to each sensor data, and the trust levels of multiple sensor data are integrated according to the evidence combination rules. Evidence theory is used to handle conflicts between different sensor data. When conflicts occur, the final information is determined by adjusting the trust level and evidence combination rules. A series of consistency check rules are defined, such as whether the vehicle's motion trajectory conforms to physical laws and whether the status of traffic signals is reasonable. Finally, a consistency check mechanism is established to perform logical verification of the global environmental perception knowledge graph, providing a reliable environmental perception knowledge graph for the autonomous driving system, which helps the system make more accurate decisions.

[0056] The group collaborative decision-making unit 2 performs spatiotemporal joint optimization decisions based on the global environmental perception knowledge graph. It models vehicle trajectories within a 100-meter radius using a Monte Carlo tree search framework, outputs a collision-free passage sequence, and aggregates historical passage data from the roadside MEC server to train a deep Q network model to generate a traffic flow allocation plan. The collision-free passage sequence and traffic flow allocation plan are then stored in the blockchain hash chain.

[0057] When modeling vehicle trajectories within a 100-meter radius based on the Monte Carlo tree search framework, the swarm collaborative decision-making unit 2 uses a Monte Carlo tree search algorithm based on deep reinforcement learning and risk perception:

[0058] The traffic environment within a radius of 100 meters is set as a reinforcement learning environment, and the state, action and reward function of the reinforcement learning environment are defined respectively. The state includes the vehicle's own state, such as position, speed, acceleration and heading angle, the state of surrounding vehicles, such as position, speed, acceleration, relative distance and angle, and the traffic signal state, such as the color of the traffic light and the remaining time. This information is integrated into a multidimensional vector as the state representation of the reinforcement learning environment. The action space includes the vehicle's discrete actions of acceleration, deceleration, maintaining the current speed, turning left, turning right, and maintaining the current direction, or acceleration and steering angle represented by continuous values. The reward function is designed in combination with safety, traffic efficiency and comfort. When the vehicle avoids a collision, it is given a positive reward. When it approaches the target position and the driving speed is high, it is given a positive reward. Positive rewards are given when the speed is reasonable, and negative rewards are given for frequent acceleration, deceleration or sharp turns. A risk perception module is introduced in each node of the Monte Carlo tree search. The risk perception module is based on a deep learning model. The risk perception module is used to evaluate the risks between vehicles in real time and input the risk assessment results into the decision-making process. For collision risk assessment, the vehicle's position, speed, acceleration and other information are input into the risk perception model, and the model outputs the collision risk probability. For traffic rule violation risk, the possibility of violating traffic rules is judged based on the traffic signal status and the vehicle's driving intention, and the corresponding risk assessment value is output. The risk assessment result is input into the decision-making process of reinforcement learning as additional information, and together with the state information, it affects the decision-making of the intelligent agent.

[0059] During the simulation of Monte Carlo tree search, the agent outputs the action probability distribution through the policy network of the reinforcement learning environment based on the current state and risk assessment results. The policy network of the reinforcement learning environment is a multi-layer perceptron. The input is the state and risk assessment value. According to the action probability distribution, the roulette wheel selection method is used to select an action as the action of this simulation. The output is the probability of each action, the optimal action is selected, the selected action is executed, and the next state is entered. The immediate reward is calculated according to the reward function, and the policy network and value network are updated according to the optimal action. After a simulation, the Monte Carlo reward method is used to calculate the target value of each state-action pair based on the reward sequence obtained during the simulation process. The policy gradient algorithm is used to update the parameters of the policy network so that the action probability distribution output by the policy network is more inclined to actions that can obtain high returns. The mean square error loss function is used to update the parameters of the value network so that the value network can more accurately estimate the value of each state, thereby improving traffic efficiency and safety.

[0060] When the swarm collaborative decision-making unit 2 aggregates the historical traffic data in the roadside MEC server to train the deep Q network model, it adopts the deep Q network algorithm based on federated learning and adversarial training:

[0061] A federated learning network consisting of roadside MEC servers and autonomous vehicles is established, with the roadside MEC servers and autonomous vehicles serving as nodes in the federated learning network. Each federated learning network node locally stores historical traffic data, which includes the vehicle's driving status, decision-making actions, and corresponding reward information in different traffic scenarios. A deep Q network model is trained locally. The deep Q network typically uses a multi-layer perceptron structure, with vehicle status information as input and the Q value of each action as output. During training, an experience replay mechanism is used to randomly sample small batches of data from the experience replay buffer for training to reduce the correlation between data. At the same time, target network technology is used to regularly update the target network parameters to improve training stability.

[0062] An adversarial training mechanism is introduced in the local training process of each federated learning network node to optimize the deep Q network model. In the local training process of each node, a discriminator network is introduced. The task of the discriminator network is to distinguish the decisions generated by the locally trained deep Q network model from the decisions in the real historical traffic data. The task of the generator network is to generate decisions that can deceive the discriminator. After the local training is completed, each federated learning network node uploads the model parameters to the central server. The central server aggregates the model parameters according to the federated learning aggregation algorithm. The specific steps of the aggregation algorithm are as follows: the weight of the model parameters uploaded by each node is calculated. Usually, the weight is proportional to the amount of data of the node. The model parameters of each node are weighted averaged to obtain the global model parameters. The central server sends the aggregated global model parameters to each federated learning network node. After receiving the global model parameters, each node replaces the parameters of the local model with the global model parameters to update the deep Q network model, which can make better decisions in different traffic scenarios and improve the overall performance of the autonomous driving system.

[0063] When generating traffic flow allocation plans, the swarm collaborative decision-making unit 2 uses a traffic flow allocation algorithm based on multi-agent deep reinforcement learning and game theory:

[0064] Each autonomous vehicle and roadside infrastructure is considered a sub-agent. Roadside infrastructure includes traffic lights and roadside sensor base stations. Each sub-agent has its own state and action space. For the autonomous vehicle agent, its state space may include its own position, speed, acceleration, heading angle, and the relative position and speed of surrounding vehicles. The action space may include acceleration, deceleration, left turn, right turn, and maintaining the current state. For the roadside infrastructure agent, taking traffic lights as an example, the state space may include the current light state, the remaining time for the light to switch, and the traffic flow at surrounding intersections. The sub-action space is the switching operation of the light state.

[0065] Each sub-agent is trained using a deep reinforcement learning algorithm. The sub-agent selects an action based on the current state and updates the sub-agent policy network and sub-agent value network based on the feedback from the environment. The sub-agent outputs a sub-action through the sub-agent policy network based on the current sub-state at each time step. After executing the sub-action, the sub-environment will give a sub-reward feedback, and the sub-agent will enter the next state. The sub-agent stores the four-tuple of these sub-states-sub-actions-sub-rewards-next-states in the experience replay buffer. During training, small batches of data are randomly sampled from the experience replay buffer to calculate the loss function of the sub-agent policy network and the sub-agent value network. For the sub-agent policy network, the goal is to maximize the cumulative reward. For the sub-agent value network, the goal is to minimize the error between the predicted value and the actual value. The parameters of the sub-agent policy network and the sub-agent value network are updated using the gradient descent algorithm.

[0066] In a multi-agent system, game theory methods are used to optimize the strategies of the agent strategy network and generate the final traffic flow allocation plan. The multi-agent system is regarded as a game, and the strategy set of each sub-agent is its sub-action space. The benefit function of each sub-agent is defined according to its goal in the traffic system. An iterative algorithm is used to solve the Nash equilibrium. In each iteration, each sub-agent updates its own strategy according to the current strategies of other sub-agents to maximize its own benefits. This process is repeated until the Nash equilibrium state is reached, that is, the strategy of each sub-agent no longer changes. When the Nash equilibrium is reached, the strategy combination of each sub-agent is the final traffic flow allocation plan. The generated traffic flow allocation plan can better balance the interests of each agent, improve the operating efficiency of the entire traffic system, and reduce congestion and delays.

[0067] When storing collision-free passage sequences and traffic flow allocation plans through the blockchain hash chain, the group collaborative decision-making unit 2 adopts a blockchain storage method based on homomorphic encryption and zero-knowledge proof:

[0068] The collision-free passage sequence and traffic flow allocation scheme are homomorphically encrypted, and the encrypted data is verified using zero-knowledge proof technology before being stored in the blockchain. Verification conditions are defined, for example, to verify whether the collision-free passage sequence complies with traffic regulations and whether the traffic flow allocation scheme is within reasonable resource constraints. The group collaborative decision-making unit 2 uses the zero-knowledge proof protocol to generate a proof to prove that the encrypted data meets the pre-defined verification conditions. The blockchain node uses the verification key to verify the proof. If the verification passes, the data is considered legal and complete. The data that has been homomorphically encrypted and verified by zero-knowledge proof is hashed to generate a hash value. The hash value is added to a new block of the blockchain as a transaction record. Each block contains the hash value of the previous block to form a hash chain, and the hash value is stored in the block of the blockchain, realizing fine-grained authorized access and sharing of data and improving the controllability and security of the data.

[0069] The dynamic control optimization unit 3 uses nonlinear model predictive control to plan the acceleration curve based on the collision-free passage sequence and traffic flow distribution plan, and integrates sliding mode control and fuzzy logic algorithm to control the lateral path tracking error.

[0070] The method of the dynamic control optimization unit 3 using nonlinear model predictive control to plan the acceleration curve based on the collision-free passage sequence and traffic flow distribution plan is as follows:

[0071] Collect data on the vehicle under various driving conditions, including the current state of the vehicle, such as position, speed, acceleration and heading angle, input control quantities, such as accelerator pedal opening, brake pedal force and the vehicle state at the next moment, establish a vehicle dynamics prediction model based on a neural network, and predict the vehicle state based on the vehicle dynamics prediction model. A multi-layer perceptron is used as the basic structure of the neural network. The number of nodes in the input layer is determined according to the dimension of the input data, including the current vehicle state and input control quantity. Each hidden layer contains a certain number of neurons, and nonlinearity is introduced through the activation function. The number of nodes in the output layer is the number of vehicle state variables to be predicted. The neural network is trained using the collected data, and the mean square error is used as the loss function. The weights and bias of the neural network are adjusted through the back propagation algorithm so that the model's prediction outputs the actual vehicle state. A prediction time domain is set to represent the number of future time steps to be predicted. The actual state of the current vehicle is input as the initial state into the vehicle dynamics prediction model. At each time step, according to the current state and the assumed input control quantity, the vehicle dynamics prediction model is used to predict the actual vehicle state. The mechanical prediction model predicts the vehicle state at the next time step, and this process is repeated until the vehicle state after time steps in the prediction domain is predicted. The optimization goal is set in the optimization process of the model predictive control, the desired vehicle state in the collision-free passage sequence and traffic flow distribution plan is defined, and the error between the predicted vehicle state and the desired state is calculated. One of the optimization goals is to minimize this tracking error in the prediction domain. In order to improve vehicle driving comfort and avoid drastic changes in acceleration, the optimization goal also includes minimizing the rate of change of acceleration. The two goals of minimizing tracking error and maximizing acceleration smoothness are weightedly combined to obtain a comprehensive optimization goal. Then, the optimization algorithm is used to solve the optimization problem of the optimization goal. During the optimization process, some practical constraints need to be considered, such as the upper and lower limits of acceleration and the limit of vehicle speed. When the constraints are met, the comprehensive optimization goal is minimized to obtain the optimal input control variable sequence, that is, the optimal acceleration curve, so that the vehicle can travel safely and comfortably while meeting the requirements of the collision-free passage sequence and traffic flow distribution plan.

[0072] When the dynamic control optimization unit 3 integrates sliding mode control and fuzzy logic algorithm to control the lateral path tracking error, a lateral path tracking method based on fuzzy sliding mode variable structure control and neural network compensation is adopted:

[0073] Design a sliding mode controller to calculate the control input based on the vehicle's lateral position error and heading error to control the vehicle's lateral motion to achieve the desired path. Select the vehicle's lateral position error and heading error As state variables, construct the state vector ,in is the state vector, To transpose the operator symbol, design the sliding surface , according to the arrival condition of sliding mode control Derive the control input of the sliding mode controller ;

[0074] A neural network compensator is designed. The neural network compensator takes the vehicle's lateral position error, heading error and control input as input and outputs the compensation for the sliding mode controller. The multi-layer perceptron is used as the neural network structure. The number of input layer nodes is 3, corresponding to the vehicle's lateral position error , heading error and the control input of the sliding mode controller , the hidden layer can be set to 1-2 layers, each layer contains several neurons, and the number of output layer nodes is 1, which is the compensation amount of the sliding mode controller During the actual operation of the vehicle, the lateral position error, heading error, sliding mode controller control input and corresponding system output error data under different working conditions are collected. This data is used to train the neural network. The mean square error is used as the loss function. The weights and bias of the neural network are adjusted through the back propagation algorithm, so that the neural network can learn the mapping relationship between the compensation amount and the input;

[0075] The control input and the compensation of the sliding mode controller are synthesized to obtain the final control input, which is used to control the steering system of the vehicle to perform lateral path tracking. and the output compensation of the neural network compensator Add them together to get the final control input , sending the final control input to the vehicle's steering system, driving the steering system to achieve the vehicle's lateral path tracking, improving the path tracking performance and stability, and enhancing the vehicle's handling safety.

[0076] In the present invention, the multi-source heterogeneous perception fusion unit 1 uses an algorithm based on event-driven and multi-scale feature matching to perform spatiotemporal alignment of multi-source heterogeneous data, construct and update the global environment perception knowledge graph, and the group collaborative decision-making unit 2 models vehicle trajectories based on the Monte Carlo tree search framework combined with deep reinforcement learning and risk perception, adopts federated learning and adversarial training to train the deep Q network model, uses multi-agent deep reinforcement learning and game theory to generate traffic flow distribution plans, and uses homomorphic encryption and zero-knowledge proof to store data. The dynamic control optimization unit 3 predicts and controls the acceleration curve of the planned control through a nonlinear model, and integrates sliding mode control and fuzzy logic algorithm to control the lateral path tracking error, thereby improving the safety of autonomous driving.

[0077] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An automatic driving system based on intelligent transportation, characterized in that: It includes a multi-source heterogeneous perception fusion unit (1), a group collaborative decision-making unit (2), and a dynamic control optimization unit (3); The multi-source heterogeneous perception fusion unit (1) acquires data of environmental point cloud, traffic signal phase and surrounding vehicle status through a sensor array, performs spatiotemporal alignment processing on the data, and generates a global environmental perception knowledge graph; The group collaborative decision-making unit (2) performs spatiotemporal joint optimization decision-making based on the global environment perception knowledge graph, models the vehicle trajectories within a radius of 100 meters based on the Monte Carlo tree search framework, outputs a collision-free passage sequence, aggregates historical passage data in the roadside MEC server to train a deep Q network model, generates a traffic flow allocation plan, and then stores the collision-free passage sequence and traffic flow allocation plan through the blockchain hash chain; The dynamic control optimization unit (3) adopts nonlinear model predictive control to plan acceleration curve based on collision-free passage sequence and traffic flow distribution scheme, and integrates sliding mode control and fuzzy logic algorithm to control lateral path tracking error.

2. The automatic driving system based on intelligent transportation according to claim 1, characterized in that: The multi-source heterogeneous perception fusion unit (1) adopts an adaptive spatiotemporal alignment algorithm based on event-driven and multi-scale feature matching when performing spatiotemporal alignment processing on multimodal data: Set specific event trigger conditions for different sensors. When an event is triggered, record the data of each sensor at that moment as keyframe data, process the keyframe data, and extract the features of the environmental point cloud, traffic signal phase, and surrounding vehicle status data respectively; A feature matching network is used to match features. Through training, the network learns the mapping relationship between different modal features. Then, the prediction and update steps of the Kalman filter are used, combined with the sensor's measurement error covariance matrix and the result of feature matching, to iteratively correct the temporal and spatial deviations of the environmental point cloud, traffic signal phase, and surrounding vehicle status data.

3. The automatic driving system based on intelligent transportation according to claim 2, characterized in that: When the multi-source heterogeneous perception fusion unit (1) generates a global environment perception knowledge graph, a knowledge graph construction method based on a graph attention network and semantic reasoning is adopted: Represent the spatiotemporally aligned data as a graph structure, and assign an initial feature vector to each node and edge in the graph structure. The initial feature vector contains features extracted from the environmental point cloud, traffic signal phase, and surrounding vehicle status data. A graph attention network is used to extract and learn the initial feature vectors of the graph structure. By stacking multiple layers of graph attention networks, the features of nodes and edges are fused with information at different levels. Pre-defined semantic rules and ontology knowledge base are then used to perform semantic reasoning on the features that fuse information at different levels. Through semantic reasoning, new relationships and attributes are added to the global environment perception knowledge graph to improve the content of the global environment perception knowledge graph.

4. The automatic driving system based on intelligent transportation according to claim 3, characterized in that: When the multi-source heterogeneous perception fusion unit (1) performs real-time updating and error correction on the generated global environment perception knowledge graph, a method based on incremental graph neural network and evidence theory is adopted: When new data is input, an incremental graph neural network algorithm is used to update the global environment perception knowledge graph. During the update process, an incremental training method is used to adjust the parameters of the incremental graph neural network according to the new data. Evidence theory is used to deal with conflicts between different sensor data, and finally a consistency check mechanism is established to perform logical verification on the global environment perception knowledge graph.

5. The automatic driving system based on intelligent transportation according to claim 1, characterized in that: The group collaborative decision-making unit (2) uses a Monte Carlo tree search algorithm based on deep reinforcement learning and risk perception when modeling vehicle trajectories within a radius of 100 meters based on the Monte Carlo tree search framework: The traffic environment within a 100-meter radius is set as a reinforcement learning environment. The state, action, and reward functions of the reinforcement learning environment are defined separately. A risk perception module is introduced at each node of the Monte Carlo tree search. The risk perception module is used to assess the risk between vehicles in real time and input the risk assessment results into the decision-making process. During the simulation of Monte Carlo tree search, the agent outputs the action probability distribution through the policy network of the reinforcement learning environment based on the current state and risk assessment results, selects the optimal action, and updates the policy network and value network based on the optimal action.

6. The automatic driving system based on intelligent transportation according to claim 5, characterized in that: The group collaborative decision-making unit (2) uses a deep Q network algorithm based on federated learning and adversarial training when training the deep Q network model based on the historical traffic data in the aggregated roadside MEC server: Establish a federated learning network consisting of roadside MEC servers and autonomous vehicles. Each federated learning network node locally stores historical traffic data and trains a deep Q network model locally. An adversarial training mechanism is introduced during the local training process of each federated learning network node to optimize the deep Q network model. After local training, each federated learning network node uploads the model parameters to the central server. The central server aggregates the model parameters according to the federated learning aggregation algorithm to obtain the global model parameters, and then sends the global model parameters to each federated learning network node to update the deep Q network model.

7. The automatic driving system based on intelligent transportation according to claim 6, characterized in that: The group collaborative decision-making unit (2) adopts a traffic flow allocation algorithm based on multi-agent deep reinforcement learning and game theory when generating a traffic flow allocation plan: Each autonomous vehicle and roadside infrastructure is considered as a sub-agent, each with its own state and action space; Each sub-agent is trained using a deep reinforcement learning algorithm. The agent selects sub-actions based on the current sub-state and updates the sub-agent policy network and sub-agent value network based on feedback from the environment. Game theory methods are used in multi-agent systems to optimize the strategies of sub-agent strategy networks and generate the final traffic flow allocation plan.

8. The automatic driving system based on intelligent transportation according to claim 7, characterized in that: The group collaborative decision-making unit (2) adopts a blockchain storage method based on homomorphic encryption and zero-knowledge proof when storing the collision-free passage sequence and traffic flow allocation plan through the blockchain hash chain: The collision-free passage sequence and traffic flow allocation plan are homomorphically encrypted, and the encrypted data is verified using zero-knowledge proof technology before being stored in the blockchain. The data that has been homomorphically encrypted and verified by zero-knowledge proof is hashed to generate a hash value, which is then stored in the block of the blockchain.

9. The automatic driving system based on intelligent transportation according to claim 1, characterized in that: The method of the dynamic control optimization unit (3) using a nonlinear model to predict and control the acceleration curve based on the collision-free traffic sequence and traffic flow distribution scheme is as follows: A vehicle dynamics prediction model based on neural network is established, and the vehicle state is predicted according to the vehicle dynamics prediction model. The optimization target is set in the optimization process of model predictive control, and then the optimization algorithm is used to solve the optimization problem of the optimization target to obtain the optimal acceleration curve.

10. The automatic driving system based on intelligent transportation according to claim 9, characterized in that: When the dynamic control optimization unit (3) integrates the sliding mode control and the fuzzy logic algorithm to control the lateral path tracking error, a lateral path tracking method based on fuzzy sliding mode variable structure control and neural network compensation is adopted: Design a sliding mode controller that calculates control inputs based on the vehicle's lateral position error and heading error to control the vehicle's lateral motion to achieve the desired path. Design a neural network compensator that takes the vehicle's lateral position error, heading error, and control input as input and outputs a compensation for the sliding mode controller. The control input and the compensation amount of the sliding mode controller are synthesized to obtain the final control input, which is used to control the vehicle's steering system to perform lateral path tracking.

Images