Road cloud integrated traffic control method for automatic driving lane changing

Abstract

The application discloses a kind of for automatic driving lane changing car road cloud integration traffic control method, comprising the following steps: S1. the multi-modal fusion of preprocessed multi-source sensor data is carried out, and environment state parameter is generated;S2. based on environment state parameter, utilize traffic flow prediction model to predict future traffic scene, risk assessment is carried out to traffic flow prediction result, and risk map containing space-time information is generated;S3. according to current traffic condition, vehicle characteristics and task demand, the vehicle driving on highway is adaptively role assigned, and local collaboration network is established based on the role assigned to each vehicle;S4. under the framework of optimized traffic control strategy, current traffic condition data and traffic flow prediction result are handled using distributed negotiation algorithm based on auction mechanism, generate multiple candidate lane changing schemes and evaluate, determine optimal lane changing strategy;S5. according to optimal lane changing strategy, coordinate each vehicle driving on highway to execute lane changing operation.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic management technology, and more specifically, to a vehicle-road-cloud integrated traffic management method for lane changing in autonomous driving. Background Technology

[0002] Vehicle-road-cloud integration technology, as an important development direction of intelligent transportation systems at present, is based on vehicle-to-everything (V2X) technology to achieve seamless connection and collaboration between vehicles, road infrastructure, and cloud computing platforms. In highway management, this technology can comprehensively perceive the traffic environment, share traffic information in real time, and collaboratively decide on traffic control strategies, improving highway traffic efficiency and reducing the risk of traffic accidents. However, the currently used V2X systems still have certain limitations in handling complex lane-changing scenarios, especially in the interaction between dedicated autonomous driving lanes and ordinary lanes.

[0003] The establishment of dedicated lanes for autonomous driving on highways offers a new approach to improving highway traffic efficiency, but it also presents new challenges for lane-changing management. Especially in complex scenarios such as near highway exits, autonomous vehicles need to change from the innermost dedicated lane to the outermost exit lane. This process involves multiple lane-changing operations, requiring consideration of safety, efficiency, and the impact on other vehicles.

[0004] However, current lane-changing control methods are ineffective in handling such complex lane-changing scenarios. Existing lane-changing control methods have the following limitations:

[0005] a) Lack of consideration for overall traffic flow, often focusing only on the optimization of individual vehicles or local areas.

[0006] b) It is difficult to adapt to complex dynamic environments, especially in situations with high-density traffic flow and mixed traffic (including autonomous vehicles and human-driven vehicles).

[0007] c) There are shortcomings in multi-vehicle coordination, making it difficult to achieve coordinated lane changing for large numbers of vehicles.

[0008] d) Limited ability to predict future traffic conditions makes it difficult to make forward-looking lane-changing decisions.

[0009] Therefore, there is an urgent need to propose a vehicle-road-cloud integrated traffic management method for lane changing in dedicated lanes for autonomous driving on highways. This method should fully apply multi-agent reinforcement learning to traffic management to achieve precise control over the safe and efficient change of autonomous vehicles from dedicated lanes to ordinary lanes to exit the highway. Summary of the Invention

[0010] This invention aims to solve the above problems and provides a vehicle-road-cloud integrated traffic management method for lane changing in autonomous driving. Based on vehicle-road-cloud integration and hierarchical reinforcement learning, it controls autonomous vehicles to safely and efficiently switch from dedicated autonomous driving lanes to ordinary lanes so as to safely exit the highway. This realizes intelligent lane changing control for autonomous vehicles traveling between dedicated autonomous driving lanes and ordinary lanes on highways.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] The vehicle-road-cloud integrated traffic management method for lane changing in autonomous driving includes the following steps:

[0013] S1. Collect multi-source sensor data, preprocess the multi-source sensor data, and input the preprocessed multi-source sensor data into a pre-established deep learning model for multimodal fusion to generate environmental state parameters;

[0014] S2. Based on the environmental state parameters, use a traffic flow prediction model to predict future traffic scenarios to obtain traffic flow prediction results, and use a risk assessment network model to conduct risk assessment on the traffic flow prediction results. Based on the traffic flow prediction results and the risk assessment results corresponding to the traffic flow prediction results, generate a risk map containing spatiotemporal information.

[0015] S3. Based on the current traffic conditions, vehicle characteristics and task requirements, adaptive role assignment is performed on vehicles traveling on the highway, dividing them into pathfinders, coordinators and executors, and establishing a local collaborative network based on the roles assigned to each vehicle.

[0016] S4. Under the framework of the optimized traffic control strategy, the current traffic condition data and traffic flow prediction results are processed using a distributed negotiation algorithm based on the auction mechanism to generate multiple candidate lane-changing schemes and evaluate them to determine the optimal lane-changing strategy. The optimized traffic control strategy is a hierarchical optimized traffic control strategy based on the hierarchical reinforcement learning method, including a global traffic flow strategy, a local area lane-changing coordination strategy, and an individual vehicle lane-changing strategy.

[0017] S5. Coordinate the lane-changing operations of all vehicles traveling on the highway according to the optimal lane-changing strategy.

[0018] Preferably, in step S1, the pre-established deep learning model is constructed using a multimodal fusion network based on an attention mechanism, including an input layer, a self-attention layer, and an output layer. The self-attention layer includes a self-attention layer for single-modal data processing and a cross-attention layer for interactive processing of information from different modalities.

[0019] The step of inputting the preprocessed multi-source sensor data into a pre-established deep learning model for multimodal fusion includes:

[0020] The input layer receives radar data, image data, and location information.

[0021] The self-attention layer processes the information of each single mode to obtain the single mode feature data corresponding to each single mode information, including: radar feature information, image feature information and position coding conversion information;

[0022] By using a cross-attention layer to exchange information between different single modalities, modal interaction feature data is obtained.

[0023] The single-modal feature data and the modal interaction feature data are processed based on a feedforward network to generate fused features, thereby obtaining environmental state parameters.

[0024] The environmental state parameters include traffic flow parameters, individual vehicle state parameters, and road environment parameters. The traffic flow parameters include the average speed and density of vehicles traveling on the highway. The individual vehicle state parameters include the position, speed, and acceleration of vehicles traveling on the highway. The road environment parameters include lane lines and curvature.

[0025] Preferably, the traffic flow prediction model is an RNN-VAE model constructed based on a combination of recurrent neural networks and variational autoencoders. The specific steps of using the traffic flow prediction model to predict future traffic scenarios and obtain traffic flow prediction results include:

[0026] The environmental state parameters are preprocessed to obtain the input data for the traffic flow prediction model;

[0027] The input data is processed using an RNN encoder to capture the temporal features of the input data, so as to map the input data to the latent variable space;

[0028] The input data representation in the latent variable space is decoded using an RNN decoder to generate future scenario predictions to obtain the traffic flow prediction results. Randomness is introduced into the latent variable space. The generation of future scenario predictions includes using variational inference of VAE and Monte Carlo random sampling to generate multiple possible future scenarios corresponding to different time points. The traffic flow prediction results include traffic flow prediction data corresponding to each possible future scenario.

[0029] Preferably, the risk assessment of the future traffic scenario using a risk assessment network model includes:

[0030] Using a Bayesian network that includes vehicle state nodes, environment state nodes, and vehicle interaction state nodes, the collision risk index and boundary crossing risk index of the future traffic scenario are calculated.

[0031] and / or

[0032] Based on the traffic flow prediction results and the corresponding risk assessment results, a risk map containing spatiotemporal information is generated, including:

[0033] The risk assessment results of the traffic flow prediction are mapped to a spatiotemporal coordinate system to form discrete risk points; the kernel density estimation method is used to smooth the discrete risk points into a continuous risk field, and the risk level is divided into risk levels by an adaptive threshold for the risk value distribution of the continuous risk field.

[0034] Acquire continuous risk field data at multiple time steps, and integrate the continuous risk field data at multiple time steps into a dynamic risk prediction map, which is used to represent the distribution and change of risk over time.

[0035] Preferably, step S3 specifically includes:

[0036] A vehicle status assessment system is established based on vehicle performance parameters, equipment status, and historical lane-changing data. A vehicle capability model is then established based on the vehicle status assessment system and environmental status assessment indicators.

[0037] Based on the vehicle capability model, the NSGA-III multi-objective optimization algorithm is used to assign roles to each driving vehicle, dividing the vehicles driving on the highway into pathfinders responsible for detecting the road conditions ahead, coordinators responsible for coordinating vehicles in the area, and executors of lane-changing operations.

[0038] The vehicle performance parameters include maximum acceleration, braking capacity, and steering response time; the equipment status includes sensor integrity data and communication quality evaluation data; the historical lane-changing data includes lane-changing success rate and cooperation effect evaluation data; and the environmental status assessment indicators include local traffic density, average vehicle speed, and weather influence factors.

[0039] Preferably, the step of assigning roles to each driving vehicle based on the vehicle capability model using the NSGA-III multi-objective optimization algorithm includes:

[0040] With lane-changing time minimization, risk minimization, and resource allocation balance as optimization objectives, and the allocation quantity of various roles, vehicle performance requirements, and spatial distribution requirements as constraints, the NSGA-III multi-objective optimization algorithm is used to process the optimization objectives and constraints. By adjusting the weights adaptively to balance the optimization objectives, Pareto optimal solution set is obtained through optimization calculation, and the initial role allocation scheme is selected from it.

[0041] The system monitors the performance of tasks in real time, iterates the initial task allocation scheme, and dynamically adjusts the previous task allocation scheme based on the monitoring results of task performance.

[0042] Preferably, the global traffic flow strategy is determined by using a graph neural network combined with a time series prediction model. The regional traffic state is updated based on the graph neural network. The graph neural network includes a three-layer graph convolutional structure. The node features of each layer of the graph convolutional structure include traffic flow parameters and regional state, and the edge features are used to characterize the intensity of vehicle flow interaction. The optimization objective of the graph neural network is to maximize regional traffic efficiency, maintain traffic flow balance, and maintain system stability.

[0043] The lane-changing coordination strategy for the local area includes: updating a pre-built multi-agent reinforcement learning model using a policy gradient method. The state space of the multi-agent reinforcement learning model includes local traffic conditions and vehicle distribution information, the action space includes lane-changing timing and spatial allocation decisions, and the reward function is a weighted function that includes lane-changing efficiency reward factors, safety reward factors, and coordination reward factors.

[0044] The individual vehicle lane-changing strategy includes: using the DDPG algorithm based on the Actor-Critic framework to control each vehicle traveling on the highway. The Actor network uses a 4-layer fully connected structure to output steering angle and acceleration control values, the Critic network uses a double-Q structure to evaluate action value, and a polynomial curve generation method is used to obtain the vehicle's trajectory planning.

[0045] Preferably, the step of processing current traffic condition data and traffic flow prediction results using a distributed negotiation algorithm based on an auction mechanism to generate multiple candidate lane-changing schemes, evaluating them, and determining the optimal lane-changing strategy specifically includes:

[0046] Based on a distributed negotiation algorithm employing a two-layer auction structure negotiation system, multiple rounds of negotiation are conducted in a local cooperative network to determine a negotiated lane-changing strategy. This strategy determines the association information between vehicles participating in the lane-changing task and the lanes. The two-layer auction structure negotiation system includes an upper-layer auction structure and a lower-layer auction structure. The upper-layer auction structure is used for lane resource allocation based on a VCG mechanism, while the lower-layer auction structure determines vehicle lane-changing permissions. Road infrastructure acts as the auction organizer, and participating vehicles submit bidding strategies based on their lane-changing needs. A credit mechanism is introduced during the negotiation process to adjust the bidding weight of vehicles during the negotiation process based on their historical behavior. The vehicle lane-changing needs include the desired time window, spatial location, and yielding cost.

[0047] Based on the negotiated lane-changing strategy, and combined with the current traffic conditions and traffic flow prediction results, a decision tree model is used to generate multiple candidate lane-changing trajectories, wherein the candidate lane-changing trajectories are as follows:

[0048] The Monte Carlo tree search method is used to evaluate the multiple candidate lane-changing trajectories and select the optimal lane-changing strategy.

[0049] Preferably, the step of generating multiple candidate lane-changing schemes based on the negotiated lane-changing strategy, combined with the current traffic conditions and traffic flow prediction results, using a decision tree model includes:

[0050] Combining preset kinematic constraints and dynamic limitations, the state prediction for a preset time range is expanded forward using dynamic programming based on a pre-built decision tree model, generating multiple lane-changing trajectory clusters. The pre-built decision tree is equipped with multiple nodes for transmitting complete state-action information. The information of the nodes is determined by negotiating lane-changing strategies and combining at least one of the current traffic conditions and traffic flow prediction results.

[0051] For each lane-changing trajectory in the lane-changing trajectory cluster, a trajectory performance index is calculated, which includes safety margin, energy consumption, and execution time.

[0052] Preferably, the evaluation of the multiple candidate lane-changing schemes based on the Monte Carlo tree search method to select the optimal lane-changing strategy includes:

[0053] The candidate lane-changing trajectories are input into a Monte Carlo search tree. The nodes of the extended Monte Carlo tree are selected using the UCB1 criterion. The value of the nodes is evaluated through rapid simulation. After a preset number of iterations, the expected benefits of each candidate lane-changing trajectory are evaluated using evaluation indicators. The scheme with the highest expected benefit is selected as the optimal lane-changing strategy. The evaluation indicators are established based on trajectory performance indicators.

[0054] The beneficial technical effects brought about by this invention are as follows:

[0055] By adopting a multi-level vehicle-road intelligent agent architecture, the driving situation of vehicles on highways is divided into three levels: macro, meso, and micro for regional control. It covers the entire highway network at the macro level, pays special attention to specific road sections at the meso level, and focuses on the lane-changing behavior of vehicles on highways at the micro level. It takes into account both local and overall considerations and effectively balances the overall optimization and local fine control of highway traffic management.

[0056] Based on a multi-layered vehicle-road intelligent agent architecture, adaptive role allocation is performed on vehicles traveling on highways according to traffic conditions and vehicle characteristics. This fully considers the driving tasks of different types of vehicles on different types of roads, reflects the emphasis on different roles of vehicles during actual driving, realizes vehicle interaction and vehicle-road cooperation on highways, and effectively improves the lane-changing efficiency of autonomous vehicles on highways.

[0057] Meanwhile, this invention optimizes the global traffic flow strategy of highways based on a hierarchical reinforcement learning method. By focusing on the functions of different levels, the macro layer is used to be responsible for global traffic flow optimization and long-term strategy formulation, providing overall task guidance; the meso layer is used to coordinate the lane-changing behavior of vehicles in local areas, balancing efficiency and safety; and the micro layer is used to execute specific lane-changing operations, achieving smooth and safe trajectory planning. This effectively solves the problems of low efficiency and difficulty in fast convergence of the optimization process in the original global traffic flow strategy formulation method, and improves the learning efficiency of the multi-level vehicle-road intelligent agent architecture. Attached Figure Description

[0058] Figure 1 An exemplary schematic diagram of a vehicle-road-cloud integrated traffic management scenario according to an embodiment of the present disclosure is shown.

[0059] Figure 2 An exemplary embodiment of the present disclosure illustrates a vehicle-road-cloud integrated traffic management method for lane changing in autonomous driving.

[0060] Figure 3 A flowchart of a method for multimodal fusion of preprocessed multi-source sensor data according to an embodiment of the present disclosure is illustrated.

[0061] Figure 4 An exemplary flowchart illustrates a method for predicting future traffic scenarios using a traffic flow prediction model to obtain traffic flow prediction results according to an embodiment of the present disclosure.

[0062] Figure 5 The illustration schematically depicts a vehicle-road-cloud integrated traffic management method for lane changing in autonomous driving, according to some other embodiments of the present disclosure. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0064] Figure 1 An exemplary schematic diagram of a vehicle-road-cloud integrated traffic management scenario according to an embodiment of the present disclosure is shown.

[0065] like Figure 1As shown, the vehicle-road-cloud integrated traffic management scenario of this disclosure includes a roadside device 101, an in-vehicle device 102, a cloud platform 103, and a network 104. In some embodiments, the roadside device 101 is equipped with millimeter-wave radar installed on the road guardrail, which can be used to collect kinematic data such as vehicle position and speed. High-definition cameras can be deployed on streetlight poles at preset intervals to collect images of vehicles traveling on the road and perform vehicle recognition and behavior analysis based on the images; lidar can be deployed on high poles, such as monitoring poles or information display poles, to collect 3D point cloud data. The in-vehicle device 102 is equipped with an omnidirectional radar system (forward millimeter-wave, lateral ultrasonic, and rearward millimeter-wave) to collect information about surrounding vehicles. Multiple cameras are used for environmental perception, and a positioning system is deployed. The in-vehicle device 102 also includes a driving control unit for locally calculating or receiving autonomous driving kinematic trajectory decision information from the cloud and converting it into control commands readable by the vehicle chassis. The cloud platform 103 includes edge servers 1031 deployed at certain intervals, responsible for local data processing and real-time fusion; and a regional control center 1032, responsible for larger-scale coordination and optimization. The network 104 is used to realize communication connections between roadside devices 101, vehicle-mounted devices 102, and the cloud platform 103. The network 104 can be established based on network devices.

[0066] In the above scenario, multimodal traffic flow data can be collected in real time to execute the method of the present invention.

[0067] Figure 2 An exemplary embodiment of the present disclosure illustrates a vehicle-road-cloud integrated traffic management method for lane changing in autonomous driving.

[0068] like Figure 2 As shown, an integrated vehicle-road-cloud traffic management method for lane changing in autonomous driving, according to an embodiment of the present disclosure, includes operations S1 to S5.

[0069] In operation S1, multi-source sensor data is collected, the multi-source sensor data is preprocessed, and the preprocessed multi-source sensor data is input into a pre-established deep learning model for multimodal fusion to generate environmental state parameters.

[0070] Multi-source sensor data can be acquired from data acquisition devices in the scene, including roadside equipment and vehicle-mounted equipment. It should be understood that multi-source sensor data can be multimodal, meaning it can include image data, radar data, and location information data, etc. After acquiring multi-source sensor data, the acquired data can be preprocessed to obtain input data that can be processed by deep learning models.

[0071] In some embodiments, the preprocessing method may include the following steps: performing state estimation by Kalman filtering; removing outliers from the data and then performing signal denoising; and performing spatiotemporal alignment on the denoised data.

[0072] Specifically, the filtering equations include the prediction step x(k|k-1)=Ax(k-1|k-1)+Bu(k) and the update step x(k|k)=x(k|k-1)+K(k)[z(k)–Hx(k|k-1)]; outliers are removed using a 5-frame window median filter; and signal denoising is performed using db4 wavelet level 3 decomposition. The spatiotemporal alignment steps may include: time synchronization based on GPS timestamps, with a maximum allowable delay of 100ms, and a double-buffering mechanism for data stream processing. Spatially, all data is uniformly transformed to the local ENU coordinate system, considering the influence of Earth's curvature, to complete spatiotemporal alignment. The transformed data after spatiotemporal alignment is verified, checking the numerical range, continuity, and physical constraints of the transformed data. Once verification is successful, the data preprocessing process is complete.

[0073] In step S1, the pre-established deep learning model is constructed using a multimodal fusion network based on an attention mechanism. This model includes an input layer, a self-attention layer, and an output layer. The self-attention layer comprises a self-attention layer for single-modal data processing and a cross-attention layer for interactive processing of information from different modalities.

[0074] Specifically, such as Figure 3 As shown, the method for inputting preprocessed multi-source sensor data into a pre-established deep learning model for multimodal fusion includes operations S11 to S14:

[0075] In operation S11, radar data, image data, and location information are received based on the input layer. In one example, the input layer may receive 128-dimensional radar features, 256-dimensional visual features, and 64-dimensional location codes for subsequent processing.

[0076] In operation S12, the single-modal information is processed based on the self-attention layer to obtain single-modal feature data corresponding to each single-modal information, including: radar feature information, image feature information, and position coding conversion information. In an exemplary embodiment, single-modal information can be processed by an 8-head self-attention layer (64 dimensions per head).

[0077] In step S13, information interaction is performed on each single modality through a cross-attention layer to obtain modal interaction feature data.

[0078] In step S14, the single-modal feature data and the modal interaction feature data are processed based on a feedforward network to generate fused features, thereby obtaining environmental state parameters. In an embodiment of the present invention, fused features can be generated through two layers of 1024-dimensional feedforward networks. During feature extraction, radar data is processed by CNN, image data uses ResNet50 to extract features, and location information is encoded and converted.

[0079] In embodiments of the present invention, the environmental state parameters include traffic flow parameters, individual vehicle state parameters, and road environment parameters. The traffic flow parameters include the average speed and density of vehicles traveling on the highway, the individual vehicle state parameters include the position, speed, and acceleration of vehicles traveling on the highway, and the road environment parameters include lane lines and curvature.

[0080] In operation S2, based on the environmental state parameters, a traffic flow prediction model is used to predict future traffic scenarios to obtain traffic flow prediction results, and a risk assessment network model is used to assess the risks of the traffic flow prediction results. Based on the traffic flow prediction results and the risk assessment results corresponding to the traffic flow prediction results, a risk map containing spatiotemporal information is generated.

[0081] The traffic flow prediction model is an RNN-VAE model built upon a combination of recurrent neural networks and variational autoencoders. This model employs an encoder-decoder structure. Specifically, in the VAE framework, input data is first transformed into a point (or distribution) in the latent variable space through an encoder network—the "encoding" process. Each dimension of this latent variable space can be considered a hidden feature or attribute of the original data. Since traffic flow data is typically sequential, the RNN, acting as the encoder, can capture dynamic changes over time and encode this information into the latent variable space. Then, new points are sampled from the latent variable space, and these points are transformed back into the original data space through the decoder network, attempting to reconstruct the original input or create new data instances similar to it. By learning the long-term dependencies and complex patterns in historical traffic flow sequence data, it can effectively predict traffic flow data in future scenarios. In one example, the encoder contains three stacked LSTM layers (256 units per layer) that encode the 128-dimensional input state into a mean and variance vector in a 32-dimensional latent variable space; the decoder also uses a three-layer LSTM structure to reconstruct the state sequence from the latent variables.

[0082] Among them, such as Figure 4 As shown, the traffic flow prediction model is used to predict future traffic scenarios and obtain traffic flow prediction results, specifically including operations S21 to S23.

[0083] In operation S21, the environmental state parameters are preprocessed to obtain the input data for the traffic flow prediction model. A typical data preprocessing method uses a sliding window to segment the time-series data, followed by standardization and feature dimensionality reduction to eliminate the influence of dimensions and reduce computational complexity, thus obtaining the input data. The window length of the sliding window can be set based on the lane-changing time. For example, it can be 5s, 10s, 15s, etc.

[0084] In operation S22, the input data is processed using an RNN encoder to capture the temporal features of the input data, so as to map the input data to the latent variable space.

[0085] In operation S23, the RNN decoder is used to decode the representation of the input data in the latent variable space to generate future scenario predictions to obtain the traffic flow prediction results. Randomness is introduced into the latent variable space. The generation of future scenario predictions includes using variational inference of VAE and Monte Carlo random sampling to generate multiple possible future scenarios corresponding to different time points. The traffic flow prediction results include traffic flow prediction data corresponding to each possible future scenario.

[0086] In embodiments of the present invention, by introducing randomness into the latent variable space and utilizing variational inference of VAEs and Monte Carlo random sampling to generate multiple possible future scenarios corresponding to different time points, the uncertainties in complex traffic flow scenarios can be better captured, improving the robustness and generalization ability of the model. Furthermore, it can enhance the diversity and reliability of predictions, thereby obtaining a more comprehensive and reliable risk map.

[0087] Furthermore, the risk assessment of the future traffic scenario using a risk assessment network model includes: using a Bayesian network comprising vehicle state nodes, environment state nodes, and vehicle interaction state nodes to calculate collision risk indicators and boundary crossing risk indicators for the future traffic scenario. Using Bayesian networks for risk assessment is beneficial for handling uncertainties and complex dependencies in traffic flow scenarios. When applying Bayesian networks, probability tables are statistically developed based on historical traffic flow data and optimized using expert knowledge to improve the accuracy of risk assessment. Simultaneously, a dynamic update mechanism is established; for example, the structure and parameters of the Bayesian network are dynamically adjusted based on real-time traffic flow data to adapt to new scenarios. When applying Bayesian networks, safety indicators such as collision risk and boundary crossing risk are calculated through forward inference.

[0088] In embodiments of the present invention, generating a risk map containing spatiotemporal information based on the traffic flow prediction result and the corresponding risk assessment result includes: mapping the risk assessment result of the traffic flow prediction result to a spatiotemporal coordinate system to form discrete risk points; using a kernel density estimation method to smooth the discrete risk points into a continuous risk field to reflect the spatiotemporal distribution of risk; classifying the risk value distribution of the continuous risk field into risk levels using an adaptive threshold, for example, classifying it into low risk, medium risk, and high risk; and acquiring continuous risk field data at multiple time steps, integrating the continuous risk field data at multiple time steps into a dynamic risk prediction map, which is used to represent the distribution changes of risk over time. In embodiments of the present invention, the risk map can be visualized, supporting interactive querying and analysis to support manual monitoring and decision analysis, thereby improving the user experience.

[0089] In S3, based on the current traffic conditions, vehicle characteristics, and task requirements, adaptive role assignment is performed on vehicles traveling on the highway. Vehicles traveling on the highway are divided into pathfinders, coordinators, and executors, and a local cooperative network is established based on the roles assigned to each vehicle.

[0090] In other words, by analyzing current traffic conditions, vehicle characteristics, and task requirements, including vehicle performance parameters, location information, and expected goals, and based on the analysis results, a multi-objective optimization algorithm is used to weigh factors such as efficiency, safety, and fairness to assign the most suitable role to each vehicle participating in lane changing, such as "pathfinder," "coordinator," and "executor." Based on the assigned roles, a local collaborative network can be established, laying the foundation for subsequent collaborative decision-making. Specifically, this can be achieved through the following steps:

[0091] First, a multi-dimensional status assessment is conducted. For example, a vehicle status assessment system can be established based on vehicle performance parameters, equipment status, and historical lane-changing data, and a vehicle capability model can be established based on the vehicle status assessment system and environmental status assessment indicators.

[0092] Specifically, the vehicle performance parameters include maximum acceleration, braking capacity, and steering response time; the equipment status includes sensor integrity data and communication quality evaluation data; the historical lane-changing data includes lane-changing success rate and cooperation effect evaluation data; and the environmental status assessment indicators include local traffic density, average vehicle speed, and weather influence factors.

[0093] Then, based on the vehicle capability model, the NSGA-III multi-objective optimization algorithm is used to assign roles to each vehicle. Vehicles on the highway are divided into pathfinders responsible for detecting road conditions ahead, coordinators responsible for coordinating vehicles within the area, and executors of lane-changing operations. When using the NSGA-III multi-objective optimization algorithm for role assignment, the optimization objectives are minimizing lane-changing time, minimizing risk, and achieving balanced resource allocation. The number of roles, vehicle performance requirements, and spatial distribution requirements are used as constraints. The NSGA-III multi-objective optimization algorithm processes these objectives and constraints, balancing them through adaptive weight adjustment. The Pareto optimal solution set is obtained through optimization calculation, and an initial role assignment scheme is selected from this set. Furthermore, the initial role assignment scheme is iteratively adjusted based on real-time monitoring of the role's performance, dynamically adjusting the scheme according to the monitoring results.

[0094] In an exemplary embodiment, the NSGA-III multi-objective optimization algorithm is used. The optimization objectives include minimizing the total time for all vehicles to complete lane changes; minimizing the total risk of all vehicles during the lane change process; maximizing the balance of resource allocation, and avoiding some vehicles from taking on too many tasks.

[0095] During the optimization process, the following constraints must be met: Role allocation quantity constraint: the number of each role type allocated must not exceed a preset upper limit; for example, the number of Pathfinders must not exceed 3. Vehicle performance requirement constraint: the performance of each vehicle (such as acceleration and braking distance) must meet the requirements of the assigned role. For example, Pathfinders require high acceleration performance. Spatial distribution requirement constraint: the spatial distribution of vehicles must meet preset conditions. For example, Pathfinders should be evenly distributed at the front, middle, and rear of the convoy. Specifically, constraints can be established using parameters such as the distance between the i-th vehicle and the center of the convoy, and minimum and maximum distance limits.

[0096] In the implementation of the NSGA-III multi-objective optimization algorithm, an initial population can be randomly generated, with each individual representing a role allocation scheme. The fitness value (i.e., the value of each optimization objective function) of each individual is calculated. Then, the population is non-dominated and sorted, dividing individuals into multiple frontiers. For example, the first frontier can include all schemes not dominated by other schemes, such as schemes that outperform other schemes in all three objectives: lane-changing time, risk level, and resource allocation balance. The second frontier can be schemes dominated by the first frontier schemes but not by other schemes. Reference points are used to maintain population diversity. When generating reference points, they can be evenly distributed in the objective space according to the number of optimization objectives (such as lane-changing time, risk level, and balance). Simulated binary crossover and polynomial mutation are used to generate the offspring population. Crossover operations can generate new role allocation schemes, such as combining pathfinders from different parents to explore better allocation methods. Mutation operations introduce diversity by randomly changing the genes of individuals, avoiding the algorithm from getting trapped in local optima. For example, mutation operations can randomly adjust the roles of some vehicles, such as turning a coordinator into a pathfinder, to explore new allocation possibilities. Combining the parent and offspring populations, select the next generation population. Dynamically adjust the weight coefficients w1, w2, and w3 based on the convergence of the optimization objective. For example, if the lane-changing time f1 converges slowly, increase w1; if the risk f2 converges quickly, decrease w2. Optimize using the NSGA-III algorithm to generate a Pareto optimal solution set. For instance, generate 50 non-dominated solutions, each representing a role allocation scheme. Based on actual needs, select an initial role allocation scheme from the Pareto optimal solution set. For example, the solution with the shortest lane-changing time, lowest risk, and most balanced resource allocation can be selected as the initial scheme.

[0097] In embodiments of the present invention, different roles are assigned to vehicles to address the interactions between different types of vehicles and different types of roads, reflecting different emphases in the actual tasks. For example, the pathfinder is responsible for collecting traffic information ahead, the coordinator is responsible for formulating local lane-changing strategies, and the executor is responsible for implementing the specific lane-changing operations. This allows for full utilization of the advantages of different vehicles on different types of roads and for different tasks, establishing an information-sharing and task-coordination mechanism among roles to improve overall lane-changing efficiency.

[0098] Next, execute operation S4.

[0099] In operation S4, under the framework of the optimized traffic control strategy, the current traffic condition data and traffic flow prediction results are processed using a distributed negotiation algorithm based on the auction mechanism to generate multiple candidate lane-changing schemes and evaluate them to determine the optimal lane-changing strategy. The optimized traffic control strategy is a hierarchical optimized traffic control strategy based on the hierarchical reinforcement learning method, including a global traffic flow strategy, a local area lane-changing coordination strategy, and an individual vehicle lane-changing strategy.

[0100] The global traffic flow strategy is determined by employing a graph neural network (GNN) combined with a time series prediction model. The regional traffic state is updated based on the GNN, which comprises a three-layer graph convolutional structure. The node features of each layer contain traffic flow parameters and regional state, while edge features characterize the intensity of vehicle flow interaction. The optimization objectives of the GNN are to maximize regional traffic efficiency, maintain traffic flow balance, and maintain system stability. In some embodiments, the GNN may contain three layers of graph convolutional structures (64 channels per layer). In this embodiment, the temporal dependence of traffic flow is captured using a time series prediction model. Time series prediction models such as LSTM, GRU, or Transformer can be applied. Spatial and temporal features are combined for joint prediction. Specifically, a GNN can be used to process the traffic network. LSTM or Transformer can be used to process the time series. Then, the spatial features output by the GNN are concatenated and fused with the temporal features output by the time series prediction model to perform joint prediction of spatiotemporal features.

[0101] The lane-changing coordination strategy for the local area includes: constructing an initial multi-agent reinforcement learning model. The state space of the multi-agent reinforcement learning model includes local traffic state and vehicle distribution information, the action space includes lane-changing timing and spatial allocation decisions, and the reward function is a weighted function containing lane-changing efficiency reward factors, safety reward factors, and coordination reward factors. The agent policy network is updated using a policy gradient method. It should be understood that a single vehicle can be used as an agent, and the learning efficiency of the reinforcement learning process can be improved through experience replay and priority sampling.

[0102] The individual vehicle lane-changing strategy includes controlling each vehicle traveling on the highway using a Deep Deterministic Policy Gradient (DDPG) algorithm based on the Actor-Critic framework. This algorithm combines deep learning and deterministic policy gradients, making it particularly suitable for solving continuous action space problems. In the model of this embodiment, the Actor network can be designed with a 4-layer fully connected structure to output steering angle and acceleration control values, while the Critic network uses a double-Q structure to evaluate action value. A polynomial curve generation method is also used to obtain the vehicle's trajectory planning. Using the Deep Deterministic Policy Gradient algorithm based on the Actor-Critic framework improves the reliability, efficiency, and stability of individual vehicle lane-changing trajectory planning.

[0103] In an embodiment of the present invention, a distributed negotiation algorithm based on an auction mechanism is used to process current traffic condition data and traffic flow prediction results, generate multiple candidate lane-changing schemes and evaluate them, and determine the optimal lane-changing strategy, specifically including operations S41 to S43.

[0104] In operation S41, a distributed negotiation algorithm based on a two-layer auction structure negotiation system is used to conduct multiple rounds of negotiation in the local cooperative network to determine a negotiated lane-changing strategy. This strategy is used to determine the association information between vehicles and lanes participating in the lane-changing task. The two-layer auction structure negotiation system includes an upper-layer auction structure and a lower-layer auction structure. The upper-layer auction structure is used for lane resource allocation based on the VCG mechanism. For example, it determines which vehicles and lanes participate in the lane-changing task. The lower-layer auction structure determines vehicle lane-changing permissions. For example, a lane-changing task can be decomposed into multiple sub-tasks, and the correspondence between vehicles, lanes, and sub-tasks can be determined through the lower-layer auction structure.

[0105] In this process, road infrastructure serves as the auction organizer. Participating vehicles submit bidding strategies based on lane-changing needs. A credit mechanism is introduced during the negotiation process to adjust the bidding weight of vehicles during the negotiation based on their historical behavior. By recording the historical task execution data of vehicles (such as lane-changing success rate, role allocation in lane-changing tasks, etc.), their subsequent bidding weight is affected.

[0106] The lane-changing requirement includes the expected time window, spatial location, and yielding cost. The expected time window is the time range within which a vehicle can complete a lane-changing operation, determined based on parameters such as lane-changing start time, lane-changing end time, and time margin. Spatial location refers to the spatial range in which the vehicle is located during the lane-changing process, determined based on parameters such as the vehicle's current position, target position, lane-changing distance, and safe clearance. Yielding cost represents the impact of a lane-changing operation on other vehicles, determined by parameters such as yielding time, yielding distance, yielding risk, and yielding priority.

[0107] In operation S42, based on the negotiated lane-changing strategy, and combined with the current traffic conditions and traffic flow prediction results, multiple candidate lane-changing trajectories are generated using a decision tree model.

[0108] Specifically, combining preset kinematic constraints and dynamic limitations, a pre-constructed decision tree model is used to dynamically predict the state over a preset time range, generating multiple lane-changing trajectory clusters. The pre-constructed decision tree has multiple nodes for transmitting complete state-action information. The information of these nodes is determined using a negotiated lane-changing strategy, combined with at least one of the current traffic conditions and traffic flow prediction results. Further, trajectory performance indicators are calculated for each lane-changing trajectory in the cluster, including safety margin, energy consumption, and execution time.

[0109] In some embodiments, the decision tree model employs a hierarchical tree structure, where each node represents the complete state and action information of the vehicle at a specific moment. The state information can include vehicle state information and environmental state information. Vehicle state information can include the vehicle's position, lateral and longitudinal coordinates in the current lane, vehicle speed, vehicle acceleration, and vehicle yaw rate, etc. Environmental state information can include the distance between the preceding and following vehicles in the current lane, and traffic flow prediction results for the target lane, such as the speed and density of vehicles in the target lane. Action information can include: steering wheel angle, accelerator pedal opening, brake pedal opening, etc. Each parent node generates child nodes based on a preset set of actions. The state of the child nodes is updated through a vehicle kinematics model. For example, based on the current vehicle speed and steering wheel angle, the vehicle's position and speed at the next time step are calculated.

[0110] During the generation of trajectory clusters, starting from the current moment, state predictions are performed forward over a preset time range (e.g., any value between 3 and 12 seconds) based on the decision tree model. Each layer of nodes filters feasible child nodes based on the vehicle's kinematic equations and constraints. For example, nodes exceeding road boundaries or less than a safe threshold distance to obstacles are removed. The complete path from the root node to a leaf node constitutes a candidate lane-changing trajectory, and all feasible paths form a lane-changing trajectory cluster. To reduce computational cost, pruning strategies can be employed. For example, only the optimal 5% of nodes are retained at each layer, and the remaining nodes are removed.

[0111] Trajectory performance metrics include safety margin, energy consumption, and execution time. Safety margin can be determined based on metrics such as minimum safe distance and time margin. Energy consumption can be determined based on parameters such as the rate of change of acceleration and lateral offset.

[0112] After obtaining multiple lane-changing trajectory clusters, operation S43 can be executed.

[0113] In operation S43, the multiple candidate lane-changing trajectories are evaluated based on the Monte Carlo tree search method to select the optimal lane-changing strategy.

[0114] For example, the candidate lane-changing trajectories are used as initial inputs to a Monte Carlo search tree. The starting point of each trajectory is a child node of the root node. Each node contains state information and action information, where the state information includes vehicle state information and environmental state information. The UCB1 criterion is used to select nodes for expanding the Monte Carlo tree. For example, starting from the root node, the child node with the highest UCB1 value is recursively selected until a node that is not fully expanded is reached. Starting from the selected node, a fast random simulation is performed to simulate the vehicle's behavior during the lane-changing process. During the simulation, actions such as steering wheel angle and accelerator pedal opening are randomly selected, and the vehicle state is updated. The node value is evaluated through fast simulation. The evaluation index is based on trajectory performance indicators, which may include, but are not limited to: safety margin, such as the minimum distance between the vehicle and surrounding vehicles during the simulation; energy consumption, such as the rate of change of acceleration and lateral offset during the simulation; execution time, such as the total time spent on the lane-changing operation during the simulation. During simulation, the selection, expansion, simulation, and backtracking steps can be repeated until a preset number of iterations is reached. Each iteration updates the node's statistics (such as cumulative reward Q(v) and number of visits N(v)). Based on the node's cumulative reward and number of visits, the expected return for each candidate lane-changing trajectory is calculated. After a preset number of iterations, the expected return of each candidate lane-changing trajectory is evaluated using assessment metrics. For example, the assessment metrics are weighted, and the scheme with the highest expected return is selected as the optimal lane-changing strategy.

[0115] After obtaining the optimal lane-changing strategy, operation S5 can be executed.

[0116] In operation S5, lane-changing operations are coordinated among all vehicles traveling on the highway according to the optimal lane-changing strategy.

[0117] Figure 5 The illustration schematically depicts a vehicle-road-cloud integrated traffic management method for lane changing in autonomous driving, according to some other embodiments of the present disclosure.

[0118] like Figure 5 As shown, in some other embodiments of this disclosure, when coordinating vehicles to perform lane-changing operations, in addition to performing the following... Figure 2 In addition to operations S1 to S5, operations S6 to S7 can also be performed.

[0119] When operating the S6, the vehicle status and surrounding environment on the highway are monitored in real time to obtain real-time detection data.

[0120] When operating S7, real-time monitoring data is used to conduct safety assessments and the lane-changing trajectory is adjusted based on the assessment results.

[0121] Optionally, when a potential hazard or abnormal situation is detected, a preset emergency control strategy can be triggered to intervene.

[0122] Exemplary security assessment steps may include operations S71 to S73.

[0123] In operation S71, a safety monitoring model is constructed, which includes a vehicle layer, a region layer, and a system layer. The vehicle layer uses Kalman filtering and fault detection algorithms to evaluate the reliability of sensor data in real time and detect whether there are any equipment malfunctions. The region layer is equipped with a vehicle interaction safety assessment model based on improved collision time and deceleration to assess the collision risk of vehicles traveling on highways. The system layer uses the analytic hierarchy process to comprehensively evaluate the overall safety status and generate a multi-dimensional risk assessment report.

[0124] In operation S72, the safety monitoring model is used to predict and control the vehicle lane-changing operation to obtain the optimal control sequence.

[0125] In operation S73, the lane-changing trajectory is adjusted based on the optimal control sequence.

[0126] In the predictive control process, the state variables of the safety monitoring model are monitored in real time. The state variables include position coordinates, velocity vector, acceleration, and yaw rate. The inputs of the safety monitoring model are steering wheel angle and accelerator pedal opening. The constraints include vehicle dynamics equations, road boundary constraints, vehicle speed limits, and safety distance requirements. An objective function is set to balance tracking error, rate of change of control quantity, and ride comfort. A real-time iterative optimization algorithm is used to solve the objective function to determine the optimal control sequence.

[0127] In one embodiment, for the vehicle layer, the Kalman filter algorithm can be used to fuse multi-sensor data and estimate vehicle state in real time, such as position, speed, and acceleration. Through prediction and update steps, the impact of sensor noise on state estimation is reduced. In fault detection algorithms, residual analysis or machine learning methods can be used to detect sensor data anomalies, such as data loss or excessive noise, thereby assessing the reliability of sensor data in real time and detecting any equipment malfunctions.

[0128] At the regional level, the improved Time-of-Collision (TTC) model calculates the collision time between a vehicle and an obstacle ahead. Dynamic weighting factors are introduced into the model to account for the impact of vehicle acceleration and road conditions on collision risk. These dynamic weighting factors can be functions related to vehicle state and environmental conditions, ranging from 0 to 1. Smaller values ​​indicate higher collision risk. Typical dynamic weighting factors include vehicle state-related factors, including but not limited to acceleration and speed change factors. Environmental condition-related factors include, but not limited to, road condition factors and visibility factors. By introducing dynamic weighting factors, the TTC calculation results are dynamically adjusted to more accurately reflect the actual collision risk. A deceleration model is used to calculate the required deceleration for the vehicle and assess whether it can decelerate within a safe distance. Combined with vehicle dynamics constraints (such as maximum deceleration), the collision risk level can be determined.

[0129] At the system level, a hierarchical model is constructed, using the evaluation results from the vehicle and region levels as input to comprehensively assess the overall safety status. The target layer represents the overall safety status; the criterion layer includes sensor reliability, collision risk, and vehicle dynamics performance; and the solution layer comprises specific safety measures (such as deceleration, lane changing, and emergency braking). This generates a risk assessment report containing the following dimensions: sensor status (normal / abnormal); collision risk level (low / medium / high); and vehicle dynamics performance (such as acceleration, deceleration, and steering ability).

[0130] Specifically, a predictive control model can be built based on the output of the safety monitoring model to optimize vehicle lane-changing operations.

[0131] Typically, the design objective function is:

[0132] J = w1·tracking error + w2·rate of change of control quantity + w3·ride comfort

[0133] Where w1, w2, and w3 are weighting coefficients.

[0134] A real-time iterative optimization algorithm (such as quadratic programming QP) is used to solve the objective function under constraints and generate the optimal control sequence. Typical optimal control sequences may include, but are not limited to: steering wheel angle sequence: δ0, δ1, δ2, ..., δN; accelerator pedal opening sequence: u0, u1, u2, ..., uN.

[0135] Based on the optimal control sequence, the lane-changing trajectory is adjusted to ensure that the vehicle meets safety and dynamic constraints during lane changes. For example, if the collision risk level is "high," the trajectory is adjusted to reduce lateral displacement or increase the safety distance.

[0136] In embodiments of the present invention, the lane-changing trajectory is dynamically adjusted based on the latest safety monitoring results and the optimal control sequence within each control cycle. For example, if a sudden deceleration of the vehicle ahead is detected, the optimal control sequence is recalculated, and the lane-changing trajectory is adjusted accordingly.

[0137] By constructing a multi-layered safety monitoring model (vehicle layer, area layer, and system layer), comprehensive safety assessment and dynamic optimization of vehicle lane-changing operations were achieved. This significantly improved the safety, adaptability, and efficiency of vehicle lane-changing operations, providing reliable support for automatic lane changing in highway scenarios.

[0138] In embodiments of the present invention, the preset emergency control strategy is set according to a three-level emergency response mechanism, which sets emergency response mechanisms for low-risk, medium-risk, and high-risk situations respectively. When a potential hazard or abnormal situation is detected as low-risk, the low-risk emergency response mechanism is activated, the lane-changing parameters are adjusted, and the original lane-changing strategy is maintained. When a potential hazard or abnormal situation is detected as medium-risk, the local cooperative network is triggered to conduct multiple rounds of negotiation to redetermine the optimal lane-changing strategy. When a potential hazard or abnormal situation is detected as high-risk, the vehicle is controlled to immediately perform emergency braking or return to the original lane, while simultaneously sending an emergency avoidance signal to surrounding vehicles.

[0139] Meanwhile, the vehicle-road-cloud integrated traffic management system for autonomous driving lane changing described in this embodiment also includes online updating and optimization of the learning models at the macro, meso, and micro levels. During operation, the system continuously collects and stores successful and failed lane-changing data of highway vehicles, and uses incremental learning algorithms to update the learning models at each level in real time, improving their adaptability to new lane-changing situations. Furthermore, the system periodically performs batch training, testing, and effect verification on the learning models at each level, comprehensively optimizing their performance and continuously improving the control performance of the vehicle-road-cloud integrated traffic management system for autonomous driving lane changing. Specifically, this includes the following:

[0140] First, by collecting and storing successful and failed lane-changing data of vehicles on highways, a standardized case library is constructed, which includes vehicle status (position, speed, acceleration), control commands (steering wheel, accelerator, braking), environmental information (weather, road conditions, traffic flow), and lane-changing results. Features are extracted from the successful lane-changing data in the standardized case library to summarize effective strategies, and failure mode analysis is performed on the failed lane-changing data in the standardized case library to optimize control strategies.

[0141] Secondly, real-time incremental learning is performed on each layer of the learning model. Different update strategies are adopted for each layer of the learning model. Specifically, for the graph neural network used to determine the macro-control strategy, an online gradient descent method is used, adjusting the network parameters of the graph neural network based on new data added in a sliding time window. For the multi-agent reinforcement learning model used to determine the meso-level coordination strategy, the agent policy network of the multi-agent reinforcement learning model is dynamically updated based on a priority-based experience replay mechanism. For the Actor-Critic framework used to determine the micro-level execution strategy, a Bayesian optimization method is used, adaptively adjusting the control parameters of the Actor-Critic framework. In this embodiment, an elastic learning rate mechanism is introduced to dynamically adjust the parameter update step size according to data freshness and importance.

[0142] Finally, the learning models of each layer after offline training are tested and validated. First, closed-loop testing is conducted in a high-fidelity simulation environment, using the Monte Carlo method to evaluate the adaptability of each layer of the reinforcement learning model. Then, real-vehicle validation is performed in a closed test field. In this embodiment, the test scenarios set in the closed test field cover normal and extreme operating conditions. Finally, a canary release strategy is adopted, selecting specific road sections and time periods for small-scale testing, and gradually expanding the deployment scope. An emergency response mechanism is established throughout the testing process to ensure a smooth transition during the update process of the vehicle-road-cloud integrated traffic management system used for autonomous driving lane changing.

[0143] The embodiments of the present invention also propose a vehicle-road-cloud integrated traffic management system for autonomous driving lane changing, applicable to autonomous vehicles on highways, specifically for formulating lane changing strategies for autonomous vehicles on highways.

[0144] The vehicle-road-cloud integrated traffic management system for autonomous driving lane changing includes roadside equipment, vehicle-mounted equipment, cloud platform, and network equipment.

[0145] The roadside equipment includes multiple sets of millimeter-wave radars, high-definition cameras, and lidar arranged at equal intervals. The millimeter-wave radars are installed on the highway guardrails to collect kinematic data of vehicles on the highway, including vehicle position and speed. The high-definition cameras are installed on the highway streetlight poles to perform vehicle identification and behavior analysis. The lidar is installed on the highway poles to collect 3D point cloud data.

[0146] The on-board equipment, installed on vehicles traveling on highways, includes an omnidirectional radar system, multiple cameras, a positioning system, and a driving control unit. The omnidirectional radar system collects vehicle information from surrounding vehicles, including forward-facing millimeter-wave radar, lateral ultrasonic radar, and rearward millimeter-wave radar. The multiple cameras are used to perceive the highway environment. The positioning system is a combination system employing a BeiDou / GPS dual-mode receiver and an inertial navigation IMU. The driving control unit locally calculates or receives autonomous driving kinematic trajectory decision information from the cloud and converts it into control commands readable by the vehicle chassis. It is understood that the autonomous driving kinematic trajectory decision information may include optimal lane-changing strategies.

[0147] The cloud platform includes multiple equally spaced edge servers and multiple equally spaced regional control centers. The edge servers are responsible for processing and real-time fusion of local data, while the regional control centers are responsible for coordinating and optimizing regional data.

[0148] The network device is used to enable communication connections between roadside equipment, vehicle-mounted equipment, and the cloud platform.

[0149] In one example, the roadside equipment includes multiple sets of equally spaced millimeter-wave radars, high-definition cameras, and lidar. The millimeter-wave radars are installed on the highway guardrails to collect kinematic data of vehicles on the highway, including vehicle position and speed. In this embodiment, the sampling frequency of the millimeter-wave radars is 20Hz, and the interval between adjacent millimeter-wave radars is set to 500m. The high-definition cameras are installed on the highway lampposts to perform vehicle identification and behavior analysis. In this embodiment, the sampling frequency of the high-definition cameras is 30fps, and the spacing between adjacent high-definition cameras is 300m. The lidars are installed on high poles on the highway to collect 3D point cloud data. In this embodiment, the sampling frequency of the lidars is 10Hz, and the spacing between adjacent lidars is 1000m.

[0150] The vehicle-mounted equipment is installed on vehicles traveling on highways and includes an omnidirectional radar system, multiple cameras, and a positioning system. The omnidirectional radar system is used to collect vehicle information of surrounding vehicles and includes forward millimeter-wave radar, lateral ultrasonic radar, and rearward millimeter-wave radar. The multiple cameras are used to perceive the highway environment. The positioning system is a combination system using a BeiDou / GPS dual-mode receiver and an inertial navigation IMU.

[0151] The cloud platform includes multiple equally spaced edge servers and multiple equally spaced regional control centers. The distance between adjacent edge servers is set to 2km, which is used to process and fuse local data in real time. The distance between adjacent regional control centers is set to 20km, which is used to coordinate and optimize data over a larger area.

[0152] In the embodiments of this invention, the aforementioned vehicle-road-cloud integrated traffic management system for autonomous driving lane changing is employed. This system collects multi-source sensor data, achieving multi-modal data fusion and generating a high-precision environmental state representation. This provides the most fundamental data source for the vehicle-road-cloud integrated traffic management method for autonomous driving lane changing of this invention. The vehicle-road-cloud integrated traffic management system effectively solves the problem of autonomous vehicles safely and efficiently navigating from dedicated lanes to ordinary lanes and exiting the highway on highways. It fully considers the safety, efficiency, and impact on other vehicles during the lane-changing process of autonomous vehicles, achieving precise control over lane changes by autonomous vehicles on highways.

[0153] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A vehicle-road cloud integrated traffic management method for automatic driving lane changing, characterized in that, The method comprises the following steps: S1. Collecting multi-source sensor data, pre-processing the multi-source sensor data, and inputting the pre-processed multi-source sensor data into a pre-established deep learning model for multi-modal fusion to generate an environment state parameter; S2. Based on the environment state parameter, using a traffic flow prediction model to predict a future traffic scene to obtain a traffic flow prediction result, and using a risk assessment network model to assess the risk of the traffic flow prediction result, and generating a risk map containing spatio-temporal information based on the traffic flow prediction result and the risk assessment result corresponding to the traffic flow prediction result; S3. According to the current traffic condition, vehicle characteristics and task demand, self-adaptive role allocation is performed on the vehicles running on the expressway, the vehicles running on the expressway are divided into pathfinders, coordinators and executors, and a local collaboration network is established based on the roles allocated to each vehicle; S4. Under the framework of an optimized traffic control strategy, a distributed negotiation algorithm based on an auction mechanism is used to process the current traffic condition data and the traffic flow prediction result, generate multiple candidate lane changing schemes and evaluate them to determine the optimal lane changing strategy, wherein the optimized traffic control strategy is a hierarchical optimization traffic control strategy based on a hierarchical reinforcement learning method, including a global traffic flow strategy, a lane changing coordination strategy in a local area, and an individual vehicle lane changing strategy; S5. According to the optimal lane changing strategy, the lane changing operation of each vehicle running on the expressway is coordinated.

2. The vehicle-road cloud integrated traffic management and control method for automatic driving lane changing according to claim 1, characterized in that, In step S1, the pre-established deep learning model is constructed using a multi-modal fusion network based on an attention mechanism, including an input layer, a self-attention layer and an output layer, the self-attention layer includes a self-attention layer for single-modal data processing and a cross-attention layer for interactive processing of different modal information; The input of the pre-processed multi-source sensor data into the pre-established deep learning model for multi-modal fusion comprises: Receiving radar data, image data and position information based on the input layer; Processing each single-modal information based on the self-attention layer to obtain single-modal feature data corresponding to each single-modal information, including radar feature information, image feature information and position encoding conversion information; Through the cross-attention layer, the information of each single-modal information is interacted to obtain modal interaction feature data; Based on the feedforward network, the single-modal feature data and the modal interaction feature data are processed to generate fusion features to obtain the environment state parameter; The environment state parameter includes traffic flow parameters, individual vehicle state parameters and road environment parameters, wherein the traffic flow parameters include the average speed and density of vehicles running on the expressway, the individual vehicle state parameters include the position, speed and acceleration of vehicles running on the expressway, and the road environment parameters include lane lines and curvature.

3. The vehicle-road cloud integrated traffic management method for automatic driving lane changing of claim 1, wherein, The traffic flow prediction model is an RNN-VAE model constructed based on the combination of a recurrent neural network and a variational autoencoder, and the traffic flow prediction model is used to predict a future traffic scene to obtain a traffic flow prediction result, which specifically comprises: Preprocessing the environmental state parameters to obtain input data of the traffic flow prediction model; processing the input data using an RNN encoder to capture the time sequence characteristics of the input data, so as to map the input data to a latent variable space; decoding the representation of the input data in the latent variable space using an RNN decoder to generate a future scene prediction to obtain the traffic flow prediction result, wherein randomness is introduced in the latent variable space, the generating of the future scene prediction includes using variational inference and Monte Carlo random sampling of VAE to generate a plurality of possible future scenes corresponding to different time points, and the traffic flow prediction result includes traffic flow prediction data corresponding to each possible future scene.

4. The vehicle-road cloud integrated traffic management and control method for automatic driving lane changing according to claim 1 or 3, characterized in that, The risk assessment of the future traffic scene by using the risk assessment network model includes: calculating the collision risk index and the out-of-bound risk index of the future traffic scene by using a Bayesian network including a vehicle state node, an environmental state node and a vehicle interaction state node; and / or generating a risk map containing spatio-temporal information based on the traffic flow prediction result and the risk assessment result corresponding to the traffic flow prediction result includes: mapping the risk assessment result of the traffic flow prediction result to a spatio-temporal coordinate system to form discrete risk points; using a kernel density estimation method, the discrete risk points are smoothed into a continuous risk field, and the risk value distribution of the continuous risk field is divided into risk levels by using an adaptive threshold; obtaining continuous risk field data of multiple time steps, and integrating the continuous risk field data of the multiple time steps into a dynamic risk prediction map, which is used to represent the distribution change of risk over time.

5. The vehicle-road cloud integrated traffic management method for automatic driving lane changing of claim 1, wherein, The step S3 specifically includes: establishing a vehicle state evaluation system based on vehicle performance parameters, equipment states and historical lane changing data, and establishing a vehicle capability model based on the vehicle state evaluation system and environmental state evaluation indexes; based on the vehicle capability model, using NSGA-III multi-objective optimization algorithm to assign roles to each driving vehicle, and dividing the vehicles driving on the highway into path explorers responsible for detecting the road conditions in front, coordinators responsible for coordinating vehicles in the region, and executors performing lane changing operations; wherein the vehicle performance parameters include maximum acceleration, braking ability and steering response time, the equipment state includes sensor integrity data and communication quality evaluation data, the historical lane changing data includes lane changing success rate and cooperation effect evaluation data, and the environmental state evaluation indexes include local traffic density, average vehicle speed and weather influence factor.

6. The vehicle-road cloud integrated traffic management and control method for automatic driving lane changing according to claim 5, characterized in that, The role assignment of each driving vehicle based on the vehicle capability model using the NSGA-III multi-objective optimization algorithm includes: taking the minimization of lane changing time, the minimization of risk degree and the balance of resource allocation as optimization objectives, taking the number of assignments of various roles, vehicle performance requirements and spatial distribution requirements as constraint conditions, processing the optimization objectives and constraint conditions by using the NSGA-III multi-objective optimization algorithm, balancing each optimization objective by using adaptive weight adjustment, and obtaining a Pareto optimal solution set by optimization calculation, and screening an initial role assignment scheme from the Pareto optimal solution set. Real-time monitoring of the effect of the role in performing the task, iteration of the initial role allocation scheme, dynamic adjustment of the role allocation scheme in the last round based on the monitoring result of the task effect.

7. The vehicle-road cloud integrated traffic management method for automatic driving lane changing of claim 1, wherein, The global traffic flow strategy is determined by adopting a graph neural network combined with a time series prediction model, wherein the regional traffic state is updated based on the graph neural network, the graph neural network includes three layers of graph convolution structure, the node features of each layer of graph convolution structure include traffic flow parameters and regional state, and the edge features are used to represent the interaction strength of the vehicle flow, and the optimization target of the graph neural network is to maximize the regional traffic efficiency, keep the traffic flow balance and keep the system stability; The lane changing coordination strategy of the local region includes: updating the pre-constructed multi-agent reinforcement learning model by using a policy gradient method, the state space of the multi-agent reinforcement learning model includes local traffic state and vehicle distribution information, the action space includes lane changing timing and spatial allocation decision, and the reward function is a weighted function including a lane changing efficiency reward factor, a safety degree reward factor and a cooperation degree reward factor; The individual vehicle lane changing strategy includes: using a DDPG algorithm based on an Actor-Critic framework to control each vehicle driving on the highway, wherein the Actor network adopts a 4-layer fully connected structure to output a steering angle and an acceleration control amount, the Critic network adopts a double Q structure to evaluate the action value, and a polynomial curve generation method is used to obtain the trajectory planning of the vehicle.

8. The vehicle-road cloud integrated traffic management method for automatic driving lane changing of claim 1, wherein, The distributed negotiation algorithm based on the auction mechanism is used to process the current traffic condition data and the traffic flow prediction result, generate a plurality of candidate lane changing schemes and evaluate them, and determine the optimal lane changing strategy, which specifically includes: The distributed negotiation algorithm based on the double-layer auction structure negotiation system is used to perform multi-round negotiation in the local cooperative network to determine the negotiation lane changing strategy, which is used to determine the association information of the vehicles participating in the lane changing task and the lanes, wherein the double-layer auction structure negotiation system includes an upper-layer auction structure and a lower-layer auction structure, the upper-layer auction structure is used for lane resource allocation based on a VCG mechanism, and the lower-layer auction structure is used for determining vehicle lane changing permission, wherein the road infrastructure is taken as an auction host, the participating vehicles submit bidding strategies according to the lane changing demand, a credit mechanism is introduced in the negotiation process, and the credit mechanism is used to adjust the bidding weight of the vehicle in the negotiation process based on the historical behavior of the vehicle in the bidding process, wherein the vehicle lane changing demand includes an expected time window, a spatial position and a yielding cost; Based on the negotiation lane changing strategy, a plurality of candidate lane changing trajectories are generated by using a decision tree model in combination with the current traffic condition and the traffic flow prediction result, wherein the candidate lane changing trajectories are generated; The plurality of candidate lane changing trajectories are evaluated by using a Monte Carlo tree search method to screen the optimal lane changing strategy.

9. The vehicle-road cloud integrated traffic management method for automatic driving lane changing of claim 8, wherein, The plurality of candidate lane changing schemes are generated by using a decision tree model in combination with the current traffic condition and the traffic flow prediction result based on the negotiation lane changing strategy, wherein the candidate lane changing schemes are generated; The state prediction in a preset time range is developed forward based on a pre-constructed decision tree model and a dynamic programming method in combination with preset kinematic constraint conditions and dynamic limit conditions, to generate multiple lane-changing trajectory clusters, wherein the pre-constructed decision tree is provided with multiple nodes for delivering complete state-action information, and the information of the nodes is determined by using a negotiation lane-changing strategy in combination with at least one of a current traffic condition and a traffic flow prediction result; A trajectory performance index is calculated for each lane-changing trajectory in the lane-changing trajectory clusters, and the trajectory performance index includes a safety margin, energy consumption, and execution time.

10. The vehicle-road cloud integrated traffic management method for automatic driving lane changing according to claim 8 or 9, characterized in that, The evaluation of the multiple candidate lane-changing schemes by using the Monte Carlo tree search method and the selection of an optimal lane-changing strategy include: The candidate lane-changing trajectories are input into a Monte Carlo search tree, a UCB1 criterion is used to select a node of the Monte Carlo tree for expansion, a node value is evaluated through fast simulation, after a preset number of iterations of sampling, an evaluation index is used to evaluate the expected returns of each candidate lane-changing trajectory, and a scheme with the highest expected return is selected as the optimal lane-changing strategy, wherein the evaluation index is established based on the trajectory performance index.

Images