Vehicle-road-cloud dynamic cooperative automatic driving method based on space-ground integrated communication network

Abstract

The application discloses a kind of vehicle-road cloud dynamic collaborative automatic driving methods based on earth-space integrated communication network, and is connected with cloud by communication double-link stable connection car end, according to whether there is roadside equipment end to formulate vehicle-road cloud collaborative strategy, support vehicle-road cloud three-level cooperation and car cloud two-level cooperation mode;At the same time, regional cloud docking center cloud to realize wide-area cloud cloud cooperation, utilize low-orbit satellite data and historical experience optimization decision result, through memory bank and teacher student model, improve overall decision performance, finally realize the maximization of collaborative automatic driving efficiency.The application improves the ability of vehicle-road cloud cooperation in wide-area coverage, communication stability and collaborative decision making through earth-space integrated communication and collaborative mechanism, improves the safety and reliability of system, and finally realizes efficient and safe collaborative automatic driving.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a vehicle-road-cloud dynamic collaborative autonomous driving method based on an integrated space-ground communication network. Background Technology

[0002] With the rapid development of information technology and intelligent control technology, autonomous driving technology, as an important component of intelligent transportation systems, has received widespread attention and rapid development. Autonomous vehicles, through a series of technologies such as perception, decision-making, and control, achieve autonomous operation in complex traffic environments, greatly improving traffic efficiency and safety.

[0003] Currently, autonomous driving technology still faces numerous challenges in areas such as wide-area coverage, real-time data transmission, and collaborative control. Existing autonomous driving systems primarily rely on ground infrastructure and limited communication networks to achieve information exchange and collaborative decision-making between vehicles, infrastructure, and the cloud. However, these systems have limitations in terms of wide-area coverage, signal stability, and data transmission latency. The coverage of ground communication base stations is limited, making it difficult to guarantee stable communication in remote or complex terrain areas. Simultaneously, with the surge in the number of autonomous vehicles, existing communication networks face problems of insufficient bandwidth and increased latency, affecting the effectiveness of real-time collaborative control.

[0004] Low Earth Orbit (LEO) satellite communication, as an emerging communication technology, boasts advantages such as wide-area coverage, low latency, and high reliability, gradually becoming an important means of supplementing terrestrial communication networks. In recent years, with the rapid deployment of LEO satellite constellations and the continuous advancement of satellite communication technology, LEO satellites have demonstrated enormous potential in providing efficient global communication services. However, the deep integration and collaborative operation of LEO satellites with terrestrial systems have not yet fully realized their advantages, particularly in the field of autonomous driving. How to effectively integrate LEO satellites with terrestrial communication networks to achieve wide-area collaboration between space and ground remains a pressing technical challenge. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a vehicle-road-cloud dynamic cooperative autonomous driving method based on an integrated space-ground communication network. The aim is to build a more comprehensive, stable, and efficient autonomous driving support system by integrating low-Earth orbit satellites and ground infrastructure. This system not only compensates for the deficiencies of terrestrial communication networks and enhances the operational capabilities of autonomous vehicles in complex and wide-area environments, but also further improves the safety and reliability of the autonomous driving system through efficient data transmission and collaborative control.

[0006] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0007] A vehicle-road-cloud dynamic cooperative autonomous driving method based on an integrated space-ground communication network:

[0008] When the vehicle terminal starts up, it activates the network interface and accesses the regional cloud via dual communication links; the dual communication links refer to the vehicle terminal simultaneously accessing both the mobile communication network and the satellite communication network.

[0009] The regional cloud dynamically groups vehicles and determines the vehicle-road-cloud collaborative strategy based on ground infrastructure conditions. The ground infrastructure conditions refer to the presence or absence of roadside equipment. If roadside equipment exists, a three-level vehicle-road-cloud collaborative strategy is set, with the roadside equipment serving as the group decision-making center and the regional cloud providing macro-level decision guidance. If roadside equipment does not exist, a two-level vehicle-cloud collaborative strategy is set, with the regional cloud serving as the group decision-making center and latency compensation enabled during decision-making.

[0010] When determining the vehicle-road-cloud collaborative strategy, the regional cloud connects with the central cloud to achieve wide-area cloud-cloud collaboration, utilizes low-orbit satellite data and historical experience to optimize decision-making results, and improves overall decision-making performance through memory banks and teacher-student models.

[0011] Furthermore, the dual communication links are divided into a primary link and a backup link. When the vehicle terminal enters the switching buffer zone between different regional clouds, both the primary link and the backup link perform data transmission tasks. If the regional cloud of the vehicle terminal changes, the primary link begins to attempt to connect to the new regional cloud, while the backup link continues to transmit data with the old regional cloud until the primary link successfully connects to the new regional cloud. At this point, the backup link disconnects from the old regional cloud and establishes a communication connection with the new regional cloud. After the vehicle terminal leaves the switching buffer zone between different regional clouds, the backup link stops transmitting data, and only the primary link performs data transmission.

[0012] Furthermore, the regional cloud dynamically groups vehicles: when a vehicle joins or leaves, the grouping mechanism is activated immediately; when there is no change in the number of vehicles, the grouping mechanism is activated periodically. In the grouping mechanism, the regional cloud clusters regional location, roadside equipment information, historical experience, and vehicle perception confidence to obtain various collaborative groups.

[0013] Furthermore, the perception confidence of the vehicle is calculated by the regional cloud based on the vehicle's configuration and location information for the newly added vehicle; the perception confidence S of the vehicle perception =a*S base +b*S history +c*S env Where a, b, and c are the calculation weights, and S base S is the vehicle's basic perception score. history S is an experience-based scoring function. env Let S be the environmental complexity score; and S base =S camera*X+S lidar *Y+S radar *Z, S camera S lidar and S radar These are the calculation weights for vehicle-mounted cameras, LiDAR, and millimeter-wave radar, respectively, and X, Y, and Z are the number of vehicle-mounted cameras, LiDAR, and millimeter-wave radars deployed on the vehicle, respectively. M represents the historical target detection accuracy of the vehicle; S env The method for obtaining the data is as follows: The regional cloud identifies the total number of vehicles N in the region based on satellite observations, and calculates the traffic flow density by combining this with the road length L in the region. Then, normalization is performed, and the normalized traffic flow density is weighted and summed with the weather level to obtain the environmental complexity score.

[0014] Furthermore, the vehicle-road-cloud three-level collaborative strategy is as follows:

[0015] The vehicle terminal and roadside equipment detect the performance of V2X communication in the field and agree on a data fusion strategy. If the bandwidth is sufficient and the latency is low, an early fusion strategy is agreed to be adopted. If the bandwidth is tight or the latency is high, a mid-term fusion strategy is agreed to be adopted.

[0016] In the initial fusion strategy, the vehicle terminal shares its raw perception data, including LiDAR point clouds, camera images, and positioning information, with the roadside equipment. It also makes individual vehicle decisions based on its own perception data, i.e., the next driving path. After receiving data from other terminals within the group, the roadside equipment performs spatiotemporal alignment of the LiDAR point cloud data based on the positioning information of other terminals and performs initial fusion. The fused data is encoded as point cloud BEV feature A, and the camera image data is mapped and encoded as visual BEV feature A. Then, point cloud BEV feature A and visual BEV feature A are fused to form fused BEV feature A, which is uploaded to the regional cloud. The roadside equipment performs target detection and prediction based on fused BEV feature A, generating preliminary decision results, i.e., the next driving path for all vehicle terminals within the group.

[0017] In the mid-term fusion strategy, the vehicle terminal first extracts features and fuses multimodal features from its own perception data to form a single-vehicle BEV feature A. After further encoding to reduce its size, it shares the feature with the roadside equipment. At the same time, it makes a single-vehicle decision based on its own perception data, i.e., the next driving path. After receiving data from other terminals in the group, the roadside equipment performs spatiotemporal alignment of the single-vehicle BEV feature A based on the positioning information of other terminals and performs mid-term fusion to form a fused BEV feature B, which is then uploaded to the regional cloud. The roadside equipment performs target detection and prediction based on the fused BEV feature B to generate a preliminary decision result, i.e., the next driving path for all vehicle terminals in the group.

[0018] After receiving the fused BEV feature A or fused BEV feature B from the roadside equipment in the group, the regional cloud makes corrections by combining the observation data of low-orbit satellites with historical experience; then, the regional cloud performs target detection and prediction, generates regional decision results, and at the same time requests macro-level traffic flow optimization targets and decision guidance from the central cloud.

[0019] After receiving a request from the regional cloud, the central cloud generates a globally optimal traffic organization plan through large-scale traffic simulation and optimization algorithms. Then, it sends the macro-level traffic flow optimization objectives and decision guidance to the regional cloud. Based on this, the regional cloud optimizes the regional decision results and sends them back to the roadside equipment in the group through the main link, thereby issuing timing or adjustment instructions to the current roadside equipment.

[0020] The roadside equipment optimizes its initial decision based on the regional decision optimization results and sends them to the vehicle terminals within the group;

[0021] After receiving the group decision results, the vehicle terminal optimizes its own single-vehicle decision results based on them, outputs waypoints based on the final decision results, calculates speed and direction based on the vehicle kinematics model and dynamics model, and finally outputs control signals to control the vehicle.

[0022] The sufficiency or scarcity of bandwidth is determined by the number of vehicle terminals and roadside equipment in the field; while the high or low latency is determined by the available bandwidth capacity.

[0023] Furthermore, the correction process, which combines low-orbit satellite observation data with historical experience, involves the following steps: The regional cloud retrieves satellite observation data based on the geographical location information of vehicles and roads. Through precise spatiotemporal coordinate alignment, the satellite observation data is mapped onto a gridded map of the current road segment. Simultaneously, the regional cloud stores vehicle driving data, perception records, and decision-making results for the region at different times and under different environments. The regional cloud compares the current BEV feature A or fused BEV feature B with similar scenarios from historical experience to calibrate scene perception errors. The fused BEV feature A or fused BEV feature B, satellite observation data, and historical experience results are then comprehensively aligned and weighted to output a corrected environmental perception result as input for subsequent target detection and prediction.

[0024] Furthermore, the roadside equipment optimizes its initial decision-making results based on the regional decision optimization results. Specifically, it first establishes optimization objectives and sets constraints, then adopts a local optimization algorithm to perform a second solution based on the macro guidance issued by the region. At the same time, it makes local fine adjustments to the speed suggestions, lane allocation, and queue sorting of grouped vehicles. If a vehicle has an urgent priority need, it is given higher weight during the local fine-tuning process. Finally, the optimized decision-making results are organized into specific instructions or suggestions that can be executed by each vehicle in the group.

[0025] Furthermore, the vehicle-cloud two-level collaborative strategy is as follows:

[0026] After receiving the packet results and coordination strategy, the vehicle terminal begins to detect the communication performance of the main links. If the bandwidth is sufficient and the latency is low, it agrees with the regional cloud to adopt early-stage fusion. If the bandwidth is tight or the latency is high, it agrees with the regional cloud to adopt mid-stage fusion.

[0027] In the initial fusion strategy, the vehicle terminal transmits its own raw perception data, including LiDAR point cloud, camera images, and positioning information, to the regional cloud through the main link, and makes individual vehicle decisions based on its own perception data. After receiving the data from each vehicle terminal, the regional cloud performs spatiotemporal alignment of the LiDAR point cloud data according to the positioning information of each vehicle terminal and performs initial fusion. The fused data is encoded as point cloud BEV feature B, and the camera image data is mapped and encoded as visual BEV feature B. A multimodal feature fusion algorithm is used to fuse the two to form fused BEV feature C.

[0028] In the mid-term fusion strategy, the vehicle terminal first extracts features and fuses multimodal features from its own perception data to form a single-vehicle BEV feature B. After further encoding to reduce its size, it is transmitted to the regional cloud through the main link. At the same time, it makes single-vehicle decisions based on its own perception data. After receiving the data from each vehicle terminal, the regional cloud performs spatiotemporal alignment of the single-vehicle BEV feature B according to its positioning information and performs mid-term fusion to form a fused BEV feature D.

[0029] The regional cloud combines low-orbit satellite observation data and historical experience to further refine the fused BEV feature C or fused BEV feature D; then, the region performs target detection and prediction, generates regional decision results, and simultaneously requests macro-level traffic flow optimization targets and decision guidance from the central cloud.

[0030] After receiving a request from the regional cloud, the central cloud combines a broader traffic flow model with global environmental information to generate macro-level traffic flow optimization goals and more specific decision-making guidance. It then sends these macro-level goals and guidance to the regional cloud. Based on this, the regional cloud optimizes the regional decision-making results, mapping and connecting the received macro-level goals with the region's own traffic conditions, road resources, and real-time traffic flow to ensure that local execution is consistent with the global goals. A delay compensation module is introduced for further advanced optimization. Finally, based on the driving needs within each group, group decision-making results are generated and transmitted back to the corresponding vehicle terminals through the main link.

[0031] After receiving the group decision results, the vehicle terminal optimizes its own single-vehicle decision results based on them, outputs waypoints based on the final decision results, calculates speed and direction based on the vehicle kinematics model and dynamics model, and finally outputs control signals to control the vehicle.

[0032] Furthermore, the regional cloud connects to the central cloud to achieve wide-area cloud-to-cloud collaboration, specifically as follows:

[0033] The regional cloud is connected to the network backbone via fiber optic links and then accesses the central cloud.

[0034] The regional cloud will generate regional decision-making results and upload them to the central cloud. Simultaneously, it will observe the movements of each group within the region using low-orbit satellites, collect the movement trajectories of vehicle terminals and their impact on the surrounding traffic environment, and evaluate the results from three aspects: safety, regional traffic efficiency, and timeliness, to obtain a decision score; the decision score S... decision =α*S safety +β*S efficiency +γ*S time Where α, β, and γ are the calculation weights, and S safety For security score, S efficiency S scores regional traffic efficiency. time Rate the timeliness;

[0035] The regional cloud encodes the regional decision-making results and their decision scores as memory features, stores them in its own memory bank, and simultaneously uploads them to the central cloud.

[0036] After receiving the memory features of each regional cloud, the central cloud stores them in the global memory bank and performs sparsification operations periodically. Through data reconstruction and model learning, it obtains the optimal or suboptimal decision operations for each region and its internal scenarios. Based on this, a high-performance teacher big model is built and continuously updated through continuous learning.

[0037] After establishing a high-performance teacher model, the central cloud performs knowledge distillation on each regional cloud to guide the training of student models in each regional cloud.

[0038] The regional cloud quickly makes optimal or suboptimal decisions based on student mini-models and its own memory bank, and transmits them back to each group through the main link, ultimately achieving efficient and safe wide-area vehicle-road-cloud cooperative autonomous driving.

[0039] Furthermore, the regional cloud periodically performs sparsification operations on its own memory bank, enhancing the memory priority of optimal or near-optimal decisions in various scenarios through data reconstruction and model learning, thereby updating its own memory bank. When encountering similar scenarios, it prioritizes referencing similar decisions in its own memory bank. For new scenarios that have never been encountered or are uncommon, the regional cloud performs online learning or short-term reinforcement learning on the collected sensor data and vehicle dynamic behavior to make preliminary decisions. At the same time, it observes the regional decision-making results, recording the decision-making process and its execution results in the local memory bank to record and accumulate new decisions. In subsequent sparsification operations or data reconstruction, if the new scenarios and new decisions can be proven effective, they will gradually receive higher memory priority, and can be preferentially invoked when the same or similar scenarios reappear in the future.

[0040] The beneficial effects of this invention are as follows:

[0041] (1) This invention accesses the regional cloud through dual links of mobile communication network and satellite communication network, ensuring that vehicles maintain stable collaborative connection in various environments and improving the reliability and fault tolerance of communication;

[0042] (2) The regional cloud in this invention dynamically groups vehicles based on their real-time configuration, location information and perception confidence, and adopts a clustering strategy to ensure that vehicles in the collaborative group can achieve the best or second-best collaborative driving effect, thereby improving overall traffic efficiency and safety.

[0043] (3) This invention supports two strategies: vehicle-road-cloud three-level collaboration and vehicle-cloud two-level collaboration. It can automatically switch according to the presence or absence of roadside collaborative equipment, adapt to different infrastructure conditions, and enhance the adaptability and scalability of the system.

[0044] (4) This invention optimizes data transmission and fusion methods based on network bandwidth and latency through early-stage and mid-stage fusion strategies, thereby improving the utilization efficiency of sensing data and the accuracy of decision-making. At the same time, it further optimizes decision-making results by combining low-orbit satellite observation data and historical experience.

[0045] (5) This invention achieves wide-area cloud-cloud collaboration by connecting regional cloud with central cloud, accumulates and optimizes decision-making experience in various scenarios through the construction of memory bank, and improves the performance of regional cloud model by using high-performance teacher big model and knowledge distillation technology, so as to achieve fast and efficient decision optimization. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0047] Figure 2 This is a schematic diagram illustrating the specific operation process of the present invention;

[0048] Figure 3 This is a schematic diagram of the wide-area cloud collaborative knowledge distillation process of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0050] This invention proposes a dynamic cooperative autonomous driving method and system for vehicles, roads, and clouds based on an integrated space-ground communication network. It combines low-Earth orbit (LEO) satellite communication with terrestrial mobile communication to achieve stable cooperation between vehicles, roads, and clouds over a wide area. The method uses dual communication links to stably connect the vehicle and cloud, and formulates vehicle-road-cloud cooperative strategies based on ground infrastructure conditions (the presence or absence of roadside equipment), supporting both three-level and two-level vehicle-road-cloud cooperative modes. Simultaneously, regional clouds connect to the central cloud to achieve wide-area cloud-cloud cooperation. LEO satellite data and historical experience are used to optimize decision-making results, and a memory bank and teacher-student models are used to improve overall decision-making performance, ultimately maximizing the efficiency of cooperative autonomous driving. This method, through integrated space-ground communication and cooperative mechanisms, significantly improves the capabilities of vehicle-road-cloud cooperation in terms of wide-area coverage, communication stability, and cooperative decision-making, enhancing system security and reliability, and ultimately achieving efficient and safe cooperative autonomous driving with broad application prospects.

[0051] like Figure 1 , 2 As shown, the present invention provides a vehicle-road-cloud dynamic cooperative autonomous driving method based on an integrated space-ground communication network, comprising the following steps:

[0052] S1. The vehicle terminal starts up, activates the network interface, and accesses the regional cloud via dual links;

[0053] S2, the regional cloud dynamically groups vehicles and determines the vehicle-road-cloud collaborative strategy based on facility conditions;

[0054] The process of autonomous driving under the S3 vehicle-road-cloud three-level collaborative strategy;

[0055] The process of autonomous driving under the S4 vehicle-cloud two-level collaborative strategy;

[0056] S5, wide-area cloud collaboration and model building process.

[0057] To further explain, step S1 specifically includes:

[0058] S11. The vehicle terminal starts up, obtains vehicle configuration and status information, initializes on-board sensors and calibrates vehicle positioning information, opens the network interface, and searches for network links.

[0059] S12, the vehicle terminal detection network link selects the best link to simultaneously access the mobile communication network and the satellite communication network (i.e., dual communication link), and then accesses the nearest regional cloud through the dual communication link;

[0060] S13. The vehicle terminal and the regional cloud periodically detect the reachability and latency of the two communication links and automatically switch the link with the lowest latency as the primary link for data transmission. The other link is a backup link to take over the communication task when the primary link is interrupted, the latency is too high, or the regional cloud is switched.

[0061] S14. When the vehicle terminal enters the switching buffer area between different regional clouds, both the main link and the backup link perform data transmission tasks. If the regional cloud of the vehicle terminal changes, the main link starts to try to connect to the new regional cloud, while the backup link continues to transmit data with the old regional cloud until the main link successfully connects to the new regional cloud. At this time, the backup link disconnects from the old regional cloud and establishes a communication connection with the new regional cloud.

[0062] S15. After the vehicle terminal leaves the switching buffer area between different regional clouds, the backup link stops data transmission and only the main link performs data transmission.

[0063] To further explain, step S2 specifically involves:

[0064] S21, the regional cloud continuously provides communication connectivity for vehicles in the region, enabling dynamic joining and leaving of vehicles;

[0065] S22. The regional cloud performs a preliminary assessment of newly added vehicles based on vehicle configuration and location information, and calculates the vehicle's perception confidence based on historical experience.

[0066] Suppose a newly added vehicle A enters the coverage area of ​​the regional cloud. The regional cloud first obtains the vehicle's hardware configuration, assigns computational weights based on the types of onboard sensors, and derives a basic perception score. Assume the computational weights for the onboard camera, LiDAR, and millimeter-wave radar are S, respectively. camera S lidar and S radar If the number of onboard cameras, LiDAR, and millimeter-wave radars on vehicle A are X, Y, and Z respectively, then the basic perception score of vehicle A is S. base =S camera *X+S lidar *Y+S radar*Z; The region cloud queries the historical perception performance of vehicles and obtains an empirical scoring function. Assuming vehicle A's historical target detection accuracy is M and its error rate is 1-M, then vehicle A's empirical scoring function is: The regional cloud identifies the total number of vehicles N within the region based on satellite observations and obtains the road length L in that region, thereby calculating the traffic density. The data is then normalized, and combined with meteorological data to obtain the current regional weather conditions, which are then divided into several levels (0-10). The higher the weather level W, the greater the perceived disturbance. Finally, the normalized traffic density is weighted and summed with the weather level to obtain the environmental complexity score S. env This is used to measure the perception difficulty of the environment in which vehicle A is located; finally, the region cloud uses vehicle A's basic perception score S to measure the perception difficulty of the environment in which vehicle A is located. base Experience scoring function S history and the environmental complexity score S env Obtain the vehicle's perception confidence S perception =a*S base +b*S history +c*S env , where a, b, and c are the calculation weights;

[0067] S23. The regional cloud dynamically groups vehicles. When a vehicle joins or leaves, the grouping mechanism is activated immediately. When there is no change in the number of vehicles, the grouping mechanism is activated periodically to achieve the stability of the grouping and the timeliness of dynamic changes.

[0068] S24. In the grouping mechanism, the regional cloud clusters regional location, roadside equipment information, historical experience and vehicle perception confidence to obtain various collaborative groups, thereby ensuring that each collaborative group can achieve the best or second-best collaborative driving effect.

[0069] S25. The regional cloud sets the vehicle-road-cloud collaboration strategy based on the roadside equipment information within each collaboration group. If roadside collaboration equipment exists, it is set to three-level vehicle-road-cloud collaboration, with the roadside end serving as the group decision center and the regional cloud providing macro-level decision guidance. If roadside collaboration equipment is missing, it is set to two-level vehicle-cloud collaboration, with the regional cloud serving as the group decision center and delay compensation enabled during decision-making.

[0070] S26. The regional cloud sends the grouping results and coordination strategies to each vehicle terminal and roadside equipment terminal, and begins to receive grouping confirmation information. After receiving confirmation from the terminal, it enables the coordination service for the confirmed terminal and continues to wait for unconfirmed terminals. If the waiting time exceeds the threshold, the unconfirmed terminal is temporarily removed from the group and regrouped when the next round of grouping mechanism is started.

[0071] To further explain, step S3 specifically involves:

[0072] S31. Under the vehicle-road-cloud three-level collaborative strategy, the regional cloud sends the grouping results and collaborative strategy to the vehicle terminals and roadside equipment terminals in the group, and simultaneously uploads them to the central cloud.

[0073] After receiving the grouping results and coordination strategy, S32, vehicle terminals and roadside equipment terminals directly connect to other terminals within the group through field V2X communication;

[0074] S33. The vehicle terminal and roadside equipment detect the performance of V2X communication in the field and agree on a data fusion strategy. If the bandwidth is sufficient and the latency is low, an early fusion strategy is agreed upon. If the bandwidth is tight or the latency is high, a mid-term fusion strategy is agreed upon. The sufficiency or tightness of bandwidth is determined by the number of vehicle terminals and roadside equipment in the field. The more vehicle terminals and roadside equipment in the field, the tighter the bandwidth, and vice versa. The specific threshold is obtained from historical experience of the scenario. The high or low latency is determined by the idle capacity of the bandwidth. The larger the idle capacity of the bandwidth, the lower the latency, and vice versa. The specific threshold is obtained from historical experience of the scenario.

[0075] S34. In the initial fusion strategy, the vehicle terminal directly shares its original perception data, including LiDAR point cloud, camera images, and positioning information, with the roadside equipment via field V2X communication. Simultaneously, it makes individual vehicle decisions based on its own perception data, i.e., the next driving path. After receiving data from other terminals within the group, the roadside equipment performs spatiotemporal alignment of the LiDAR point cloud data based on the positioning information of other terminals and performs initial fusion. The fused data is encoded as point cloud BEV feature A, and the camera image data is mapped and encoded as visual BEV feature A. A multimodal feature fusion algorithm is used to fuse point cloud BEV feature A and visual BEV feature A to form fused BEV feature A, which is then uploaded to the regional cloud. Simultaneously, the roadside equipment performs target detection and prediction based on fused BEV feature A, generating preliminary decision results, i.e., the next driving path for all vehicle terminals within the group.

[0076] S35. In the mid-term fusion strategy, the vehicle terminal first extracts features and fuses multimodal features from its own perception data to form a single-vehicle BEV feature A. After further encoding to reduce its size, it shares the feature with the roadside equipment terminal via field V2X communication. At the same time, it makes a single-vehicle decision based on its own perception data, i.e., the next driving path. After receiving data from other terminals in the group, the roadside equipment terminal performs spatiotemporal alignment of the single-vehicle BEV feature A based on the positioning information of other terminals and performs mid-term fusion (e.g., using maxout for feature fusion, retaining the larger value at the corresponding position in multiple feature maps) to form a fused BEV feature B, which is then uploaded to the regional cloud. Simultaneously, the roadside equipment terminal performs target detection and prediction based on the fused BEV feature B to generate a preliminary decision result, i.e., the next driving path for all vehicle terminals in the group.

[0077] S36. After receiving the fused BEV features A / B from the roadside equipment in the group, the regional cloud further combines the observation data of low-orbit satellites and historical experience to make corrections. Then, the regional cloud performs target detection and prediction, generates regional decision results, and requests macro traffic flow optimization objectives (such as large-scale traffic diversion strategies across urban areas and road networks) and decision guidance (such as suggesting that a certain traffic flow be diverted from a specific exit, suggesting temporary changes to the speed limit of some road sections, or even activating emergency lanes, etc.) from the central cloud.

[0078] Specifically, the correction is made by combining low-orbit satellite observation data with historical experience. The regional cloud retrieves satellite observation data based on the geographical location information of vehicles and roads, and maps the satellite observation data to the gridded map of the current road segment (such as the BEV coordinate system) through precise spatiotemporal coordinate alignment. At the same time, the regional cloud stores vehicle driving data, perception records and decision results of the area under different times and environments (weather, time period, traffic flow). The regional cloud compares the current BEV features A / B with similar scenarios in historical experience (similar weather, road structure, traffic flow level, etc.) to calibrate the scene perception error. The fused BEV features A / B, satellite observation data and historical experience results are comprehensively aligned and weighted to output a corrected environmental perception result as the input for subsequent target detection and prediction, providing a more accurate description of the environmental state for the next step of regional decision-making.

[0079] S37. After receiving a request from the regional cloud, the central cloud makes a comprehensive decision based on traffic data and models from a larger scope (it can simultaneously dispatch information from multiple regional clouds or national-level traffic centers). Through large-scale traffic simulation and optimization algorithms, it generates globally optimal or near-optimal traffic organization schemes (such as vehicle diversion guidance, traffic light linkage across the entire area, emergency accident handling schemes, etc.). Then, it sends the macro-level traffic flow optimization goals and decision guidance to the regional cloud. Based on this, the regional cloud optimizes the regional decision results and sends them back to the roadside equipment in the group through the main link. This allows it to issue timing or adjustment instructions to the current roadside equipment, such as coordinating with signal linkage at adjacent intersections or adjacent road segments, instructing vehicles to perform diversion or maintain speed, etc.

[0080] Based on this, the regional cloud optimizes the regional decision-making results by integrating and adjusting the macro-level traffic optimization strategies issued by the central cloud with the existing local decision-making schemes of the regional cloud to ensure that the local strategies are consistent with the global objectives.

[0081] S38. The roadside equipment optimizes its own preliminary decision based on the regional decision optimization results and sends them to the vehicle terminals in the group through field V2X communication.

[0082] The roadside equipment optimizes its initial decision-making results based on the regional decision optimization results. Specifically, it first establishes an optimization objective: maximizing local traffic efficiency and safety, reducing inter-vehicle interaction conflicts, avoiding collision risks, maximizing throughput, alleviating congestion at bottlenecks, and accommodating priority strategies for special vehicles (such as ambulances and buses) without violating regional decision instructions. Then, it sets constraints, such as road speed limits, lane capacity, traffic light timing, road construction zones, or no-entry zones, as well as safety regulations and environmental restrictions. Finally, it employs local optimization algorithms (such as linear / nonlinear programming and graph search algorithms). The algorithm employs a reinforcement learning local decision-making module, which performs secondary solutions based on macro-level guidance issued by the region. It also comprehensively considers factors such as safety, efficiency, and comfort, and makes local fine-tuning of speed suggestions, lane allocation, and queue order for grouped vehicles (i.e., speed suggestions, lane allocation, and queue order are used as input parameters for the local optimization algorithm). If a vehicle has an urgent priority need, it is given higher weight during the local fine-tuning process. Finally, the optimized decision results are organized into specific instructions or suggestions that each vehicle in the group can execute (such as target lane / path, expected speed range, driving priority, or following queue information).

[0083] S39. After receiving the group decision results, the vehicle terminal optimizes its own single-vehicle decision results based on them, and then outputs waypoints based on the final decision results. It calculates speed and direction based on the vehicle kinematics model and dynamics model, and finally outputs control signals to control the vehicle.

[0084] After receiving the group decision results, the vehicle terminal optimizes its own single-vehicle decision results based on them. Specifically, it uses algorithms such as local reinforcement learning to comprehensively weigh the group decision results with the vehicle's local perception results and generate the final decision result. If a sudden situation is detected locally (such as an obstacle intruding into the lane or an emergency), the vehicle terminal can make corresponding adjustments to the group decision while adhering to the principle of safety first.

[0085] To further explain, step S4 specifically involves:

[0086] S41. Under the vehicle-cloud two-level collaboration strategy, the regional cloud sends the grouping results and collaboration strategy to the vehicle terminals in the group and uploads them to the central cloud simultaneously.

[0087] S42. After receiving the packet results and coordination strategy, the vehicle terminal begins to detect the communication performance of the main links. If the bandwidth is sufficient and the latency is low, it agrees with the regional cloud to adopt the early fusion strategy. If the bandwidth is tight or the latency is high, it agrees with the regional cloud to adopt the mid-term fusion strategy.

[0088] S43. In the early fusion strategy, the vehicle terminal transmits its own raw perception data, including LiDAR point cloud, camera image and positioning information, to the regional cloud through the main link. At the same time, it makes single-vehicle decisions based on its own perception data. After receiving the data from each vehicle terminal, the regional cloud performs spatiotemporal alignment of the LiDAR point cloud data according to the positioning information of each vehicle terminal and performs early fusion. The fused data is encoded as point cloud BEV feature B, and the camera image data is mapped and encoded as visual BEV feature B. The two are fused using a multimodal feature fusion algorithm to form fused BEV feature C.

[0089] S44. In the mid-term fusion strategy, the vehicle terminal first extracts features and fuses multimodal features from its own perception data to form a single-vehicle BEV feature B. After further encoding to reduce its size, it is transmitted to the regional cloud through the main link. At the same time, it makes single-vehicle decisions based on its own perception data. After receiving the data from each vehicle terminal, the regional cloud performs spatiotemporal alignment of the single-vehicle BEV feature B according to its positioning information and performs mid-term fusion to form a fused BEV feature D.

[0090] S45, the regional cloud, combined with the observation data of low-orbit satellites and historical experience, further corrects the fused BEV feature C / D, and then performs target detection and prediction to generate regional decision results. At the same time, it requests macro traffic flow optimization targets and decision guidance from the central cloud.

[0091] S46. After receiving a request from the regional cloud, the central cloud combines a broader traffic flow model with global environmental information to generate macro-level traffic flow optimization objectives (e.g., diversion strategies, trunk line priority passage schemes, long-distance fleet scheduling rules, etc.) and more specific decision guidance (e.g., speed limits on specific road sections, priority guidance for specific vehicle groups, restrictions on traffic flow in certain directions, etc.). The central cloud then sends the macro-level traffic flow optimization objectives and decision guidance to the regional cloud. Based on this, the regional cloud optimizes the regional decision results (mapping and connecting the received macro-level objectives with the region's own traffic conditions, road resources, real-time traffic flow, etc., to ensure that local execution is consistent with the global objectives). It also introduces the existing delay compensation module for further advanced optimization. Finally, based on the driving needs within each group, it generates group decision results and sends them back to the corresponding vehicle terminals through the main link.

[0092] Based on the driving demand within each group, group decision results are generated. Specifically, the regional cloud combines the differences in destination, current route, travel attributes (private vehicles, commercial vehicles, public transportation, etc.) and time requirements (such as estimated arrival time) of vehicles in each group to determine which vehicle groups need to be prioritized for smooth traffic flow and which vehicle groups can be diverted or temporarily wait. At the same time, algorithms such as linear / nonlinear programming, reinforcement learning, and model predictive control are used to solve the problem comprehensively and generate specific executable instructions or suggestions for each vehicle (such as target lane / route, expected speed range, driving priority or following queue information, etc.).

[0093] S47. After receiving the group decision results, the vehicle terminal optimizes its own single-vehicle decision results based on them (same as S39), generates the final decision results, outputs waypoints, and then calculates speed and direction based on the vehicle kinematics model and dynamics model, and finally outputs control signals to control the vehicle.

[0094] For further explanation, see [link to documentation]. Figure 3 Step S5 specifically includes:

[0095] S51. The regional cloud connects to the network backbone via fiber optic links, accesses the central cloud, and realizes data interaction between the cloud and external cloud through the central cloud, further improving its decision-making performance.

[0096] S52. After generating the regional decision results, the regional cloud uploads them to the central cloud in the background. At the same time, the regional cloud further observes the movement of each group in the region through low-orbit satellites, collects the movement trajectory of the vehicle terminal and its impact on the surrounding traffic environment, and evaluates the decision score from three aspects: safety, regional traffic efficiency and timeliness.

[0097] The decision score is obtained by evaluating three aspects: safety, regional traffic efficiency, and timeliness. Specifically, the following steps are taken: First, based on the collected collision risk index, frequency of rapid acceleration / braking, and number of insufficient minimum distance warnings, the data are indexed, weighted, and averaged. Then, the average is normalized to obtain the safety score S. safety Using average vehicle speed, vehicle throughput, and intersection queue length / length as core indicators, the regional traffic efficiency score S is obtained by normalizing the data by comparing it with historical averages or expected thresholds. efficiency The timeliness score S is obtained by indexing the delays in comprehensive decision-making, instruction execution, and traffic condition improvement, and then weighting and averaging these delays. time Finally, the decision score S is obtained by weighting and combining the indicators from the three dimensions. decision =α*S safety +β*S efficiency +γ*S time , where α, β, and γ are the calculation weights;

[0098] S53. The regional cloud encodes the regional decision-making results and their decision scores into memory features, stores them in its own memory bank, and simultaneously uploads them to the central cloud.

[0099] S54. The regional cloud periodically performs sparsification operations on its own memory bank. Through data reconstruction and model learning, it enhances the memory priority of optimal or suboptimal decisions in various scenarios, thereby updating its own memory bank. When encountering similar scenarios, it can prioritize referencing similar decisions in its own memory bank, improving decision performance and response speed. For new scenarios that have never been encountered or are uncommon, the regional cloud can perform online learning or short-term reinforcement learning on the collected sensor data, vehicle dynamic behavior, and other information, enabling the system to make preliminary decisions when no existing experience is available. At the same time, it observes the decision results and records the decision process and its execution results in its own memory bank to record and accumulate new experience. In subsequent sparsification operations or data reconstruction, if these new scenarios and new decisions are proven to be effective, they will gradually receive higher memory priority and can be prioritized when the same or similar scenarios reappear in the future.

[0100] S55. After receiving the memory features of each regional cloud, the central cloud stores them in the global memory bank and performs sparsification operations periodically. Through data reconstruction and model learning, it obtains the optimal or suboptimal decision operations for each region and its internal scenarios. Based on this, a high-performance teacher big model is established and continuously updated through continuous learning.

[0101] S56. After establishing a high-performance teacher model, the central cloud performs knowledge distillation on each regional cloud to guide the training of student models in each regional cloud, thereby improving the performance of the regional cloud models.

[0102] S57 and the regional cloud quickly make optimal or suboptimal decisions based on student small models and their own memory banks, and transmit them back to each group through the main link, ultimately achieving efficient and safe wide-area vehicle-road-cloud cooperative autonomous driving.

[0103] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A vehicle-road-cloud dynamic cooperative autonomous driving method based on an integrated space-ground communication network, characterized by: When the vehicle terminal starts up, it activates the network interface and accesses the regional cloud via dual communication links; the dual communication links refer to the vehicle terminal simultaneously accessing both the mobile communication network and the satellite communication network. The regional cloud dynamically groups vehicles and determines the vehicle-road-cloud collaborative strategy based on ground infrastructure conditions. The ground infrastructure conditions refer to the presence or absence of roadside equipment. If roadside equipment exists, a three-level vehicle-road-cloud collaborative strategy is set, with the roadside equipment serving as the group decision-making center and the regional cloud providing macro-level decision guidance. If roadside equipment does not exist, a two-level vehicle-cloud collaborative strategy is set, with the regional cloud serving as the group decision-making center and latency compensation enabled during decision-making. When determining the vehicle-road-cloud collaborative strategy, the regional cloud connects with the central cloud to achieve wide-area cloud-cloud collaboration, utilizes low-orbit satellite data and historical experience to optimize decision-making results, and improves overall decision-making performance through memory banks and teacher-student models. The vehicle-road-cloud three-level collaborative strategy is as follows: The vehicle terminal and roadside equipment detect the performance of V2X communication in the field and agree on a data fusion strategy. If the bandwidth is sufficient and the latency is low, an early fusion strategy is agreed to be adopted. If the bandwidth is tight or the latency is high, a mid-term fusion strategy is agreed to be adopted. In the initial fusion strategy, the vehicle terminal shares its raw perception data, including LiDAR point clouds, camera images, and positioning information, with the roadside equipment. It also makes individual vehicle decisions based on its own perception data, i.e., the next driving path. After receiving data from other terminals within the group, the roadside equipment performs spatiotemporal alignment of the LiDAR point cloud data based on the positioning information of other terminals and performs initial fusion. The fused data is encoded as point cloud BEV feature A, and the camera image data is mapped and encoded as visual BEV feature A. Then, point cloud BEV feature A and visual BEV feature A are fused to form fused BEV feature A, which is uploaded to the regional cloud. The roadside equipment performs target detection and prediction based on fused BEV feature A, generating preliminary decision results, i.e., the next driving path for all vehicle terminals within the group. In the mid-term fusion strategy, the vehicle terminal first extracts features and fuses multimodal features from its own perception data to form a single-vehicle BEV feature A. After further encoding to reduce its size, it shares the feature with the roadside equipment. At the same time, it makes a single-vehicle decision based on its own perception data, i.e., the next driving path. After receiving data from other terminals in the group, the roadside equipment performs spatiotemporal alignment of the single-vehicle BEV feature A based on the positioning information of other terminals and performs mid-term fusion to form a fused BEV feature B, which is then uploaded to the regional cloud. The roadside equipment performs target detection and prediction based on the fused BEV feature B to generate a preliminary decision result, i.e., the next driving path for all vehicle terminals in the group. After receiving the fused BEV feature A or fused BEV feature B from the roadside equipment in the group, the regional cloud makes corrections by combining the observation data of low-orbit satellites with historical experience; then, the regional cloud performs target detection and prediction, generates regional decision results, and at the same time requests macro-level traffic flow optimization targets and decision guidance from the central cloud. After receiving a request from the regional cloud, the central cloud generates a globally optimal traffic organization plan through large-scale traffic simulation and optimization algorithms. Then, it sends the macro-level traffic flow optimization objectives and decision guidance to the regional cloud. Based on this, the regional cloud optimizes the regional decision results and sends them back to the roadside equipment in the group through the main link, thereby issuing timing or adjustment instructions to the current roadside equipment. The roadside equipment optimizes its initial decision based on the regional decision optimization results and sends them to the vehicle terminals within the group; After receiving the group decision results, the vehicle terminal optimizes its own single-vehicle decision results based on them, outputs waypoints based on the final decision results, calculates speed and direction based on the vehicle kinematics model and dynamics model, and finally outputs control signals to control the vehicle. The sufficiency or scarcity of bandwidth is determined by the number of vehicle terminals and roadside equipment in the field; while the high or low latency is determined by the available bandwidth capacity.

2. The vehicle-road-cloud dynamic cooperative autonomous driving method according to claim 1, characterized in that, The dual communication links are divided into a primary link and a backup link. When the vehicle terminal enters the switching buffer zone between different regional clouds, both the primary link and the backup link perform data transmission tasks. If the regional cloud of the vehicle terminal changes, the primary link begins to attempt to connect to the new regional cloud, while the backup link continues to transmit data with the old regional cloud until the primary link successfully connects to the new regional cloud. At this point, the backup link disconnects from the old regional cloud and establishes a communication connection with the new regional cloud. After the vehicle terminal leaves the switching buffer zone between different regional clouds, the backup link stops transmitting data, and only the primary link performs data transmission.

3. The vehicle-road-cloud dynamic cooperative autonomous driving method according to claim 1, characterized in that, The regional cloud dynamically groups vehicles: when a vehicle joins or leaves, the grouping mechanism is activated immediately; when there is no change in the number of vehicles, the grouping mechanism is activated periodically. In the grouping mechanism, the regional cloud clusters regional location, roadside equipment information, historical experience, and vehicle perception confidence to obtain various collaborative groups.

4. The vehicle-road-cloud dynamic cooperative autonomous driving method according to claim 3, characterized in that, The perception confidence level of the vehicle is calculated by the regional cloud based on the vehicle's configuration and location information for the newly added vehicle; the perception confidence level of the vehicle ,in , , To calculate the weights, This is the vehicle's basic perception score. For experience-based scoring functions, The environmental complexity score; and , , and These are the calculation weights for the vehicle-mounted camera, LiDAR, and millimeter-wave radar, respectively. , and These refer to the number of vehicle-mounted cameras, lidar, and millimeter-wave radars deployed on the vehicle. M represents the historical target detection accuracy of the vehicle; The method for obtaining the data is as follows: The regional cloud identifies the total number of vehicles N in the region based on satellite observations, and calculates the traffic flow density by combining this with the road length L in the region. Then, normalization is performed, and the normalized traffic flow density is weighted and summed with the weather level to obtain the environmental complexity score.

5. The vehicle-road-cloud dynamic cooperative autonomous driving method according to claim 1, characterized in that, The correction process, which combines low-orbit satellite observation data with historical experience, involves the following steps: The regional cloud retrieves satellite observation data based on the geographical location information of vehicles and roads. Through precise spatiotemporal coordinate alignment, the satellite observation data is mapped onto a gridded map of the current road segment. Simultaneously, the regional cloud stores vehicle driving data, perception records, and decision results for the region at different times and under different environments. The regional cloud compares the current BEV feature A or fused BEV feature B with similar scenarios from historical experience to calibrate scene perception errors. The fused BEV feature A or fused BEV feature B, satellite observation data, and historical experience results are then comprehensively aligned and weighted to output a corrected environmental perception result as input for subsequent target detection and prediction.

6. The vehicle-road-cloud dynamic cooperative autonomous driving method according to claim 5, characterized in that, The roadside equipment optimizes its initial decision-making results based on the regional decision optimization results. Specifically, it first establishes optimization objectives and sets constraints, then adopts a local optimization algorithm to perform a second solution based on the macro guidance issued by the region. At the same time, it makes local fine adjustments to the speed suggestions, lane allocation, and queue order of grouped vehicles. If a vehicle has an emergency priority need, it is given higher weight during the local fine-tuning process. Finally, the optimized decision results are organized into specific instructions or suggestions that can be executed by each vehicle in the group.

7. The vehicle-road-cloud dynamic cooperative autonomous driving method according to claim 6, characterized in that, The vehicle-cloud two-level collaborative strategy is as follows: After receiving the grouping results and coordination strategy, the vehicle terminal begins to detect the communication performance of the main links. If the bandwidth is sufficient and the latency is low, it agrees with the regional cloud to adopt early-stage fusion. If the bandwidth is tight or the latency is high, it agrees with the regional cloud to adopt mid-stage fusion. In the initial fusion strategy, the vehicle terminal transmits its own raw perception data, including LiDAR point cloud, camera images, and positioning information, to the regional cloud through the main link, and makes individual vehicle decisions based on its own perception data. After receiving the data from each vehicle terminal, the regional cloud performs spatiotemporal alignment of the LiDAR point cloud data according to the positioning information of each vehicle terminal and performs initial fusion. The fused data is encoded as point cloud BEV feature B, and the camera image data is mapped and encoded as visual BEV feature B. A multimodal feature fusion algorithm is used to fuse the two to form fused BEV feature C. In the mid-term fusion strategy, the vehicle terminal first extracts features and fuses multimodal features from its own perception data to form a single-vehicle BEV feature B. After further encoding to reduce its size, it is transmitted to the regional cloud through the main link. At the same time, it makes single-vehicle decisions based on its own perception data. After receiving the data from each vehicle terminal, the regional cloud performs spatiotemporal alignment of the single-vehicle BEV feature B according to its positioning information and performs mid-term fusion to form a fused BEV feature D. The regional cloud combines low-orbit satellite observation data and historical experience to further refine the fused BEV feature C or fused BEV feature D; then, the region performs target detection and prediction, generates regional decision results, and simultaneously requests macro-level traffic flow optimization targets and decision guidance from the central cloud. After receiving a request from the regional cloud, the central cloud combines a broader traffic flow model with global environmental information to generate macro-level traffic flow optimization goals and more specific decision-making guidance. It then sends these macro-level goals and guidance to the regional cloud. Based on this, the regional cloud optimizes the regional decision-making results, mapping and connecting the received macro-level goals with the region's own traffic conditions, road resources, and real-time traffic flow to ensure that local execution is consistent with the global goals. A delay compensation module is introduced for further advanced optimization. Finally, based on the driving needs within each group, group decision-making results are generated and transmitted back to the corresponding vehicle terminals through the main link. After receiving the group decision results, the vehicle terminal optimizes its own single-vehicle decision results based on them, outputs waypoints based on the final decision results, calculates speed and direction based on the vehicle kinematics model and dynamics model, and finally outputs control signals to control the vehicle.

8. The vehicle-road-cloud dynamic cooperative autonomous driving method according to claim 1, characterized in that, The regional cloud connects to the central cloud to achieve wide-area cloud-to-cloud collaboration, specifically as follows: The regional cloud is connected to the network backbone via fiber optic links and then accesses the central cloud. The regional cloud will generate regional decision-making results and upload them to the central cloud. Simultaneously, it will observe the movements of each group within the region using low-orbit satellites, collecting the movement trajectories of vehicle terminals and their impact on the surrounding traffic environment. A decision score will be obtained by evaluating the results from three aspects: safety, regional traffic efficiency, and timeliness. ,in , , To calculate the weights, For security score, Scoring of regional traffic efficiency. Rate the timeliness; The regional cloud encodes the regional decision-making results and their decision scores as memory features, stores them in its own memory bank, and simultaneously uploads them to the central cloud. After receiving the memory features of each regional cloud, the central cloud stores them in the global memory bank and performs sparsification operations periodically. Through data reconstruction and model learning, it obtains the optimal or suboptimal decision operations for each region and its internal scenarios. Based on this, a high-performance teacher big model is built and continuously updated through continuous learning. After establishing a high-performance teacher model, the central cloud performs knowledge distillation on each regional cloud to guide the training of student models in each regional cloud. The regional cloud quickly makes optimal or suboptimal decisions based on student mini-models and its own memory bank, and transmits them back to each group through the main link, ultimately achieving efficient and safe wide-area vehicle-road-cloud cooperative autonomous driving.

9. The vehicle-road-cloud dynamic cooperative autonomous driving method according to claim 8, characterized in that, The regional cloud periodically performs sparsification operations on its own memory bank, enhancing the memory priority of optimal or suboptimal decisions in various scenarios through data reconstruction and model learning, thereby updating its own memory bank. When encountering similar scenarios, it prioritizes referencing similar decisions in its own memory bank. For new scenarios that have never been encountered or are uncommon, the regional cloud performs online learning or short-term reinforcement learning on the collected sensor data and vehicle dynamic behavior to make preliminary decisions. At the same time, it observes the regional decision results, recording the decision-making process and its execution results in the local memory bank to record and accumulate new decisions. In subsequent sparsification operations or data reconstruction, if the new scenarios and new decisions are proven to be effective, they gradually receive higher memory priority and can be preferentially invoked when the same or similar scenarios reappear in the future.

Images