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

Abstract

The invention discloses a vehicle-road cloud dynamic collaborative automatic driving method based on a space-ground integrated communication network, and the method comprises the steps: stably connecting a vehicle end and a cloud end through a communication double link, formulating a vehicle-road cloud collaborative strategy according to the existence of a roadside equipment end, and supporting a vehicle-road cloud three-stage collaborative mode and a vehicle-cloud two-stage collaborative mode; meanwhile, the regional cloud is in butt joint with the central cloud to achieve wide-area cloud-cloud collaboration, the decision result is optimized through low-orbit satellite data and historical experience, the overall decision performance is improved through a memory bank and a teacher-student model, and finally the maximization of collaborative automatic driving efficiency is achieved. According to the invention, through a space-ground integrated communication and cooperation mechanism, the capabilities of vehicle-road cloud cooperation in the aspects of wide area coverage, communication stability and cooperative decision making are significantly improved, the safety and reliability of the system are improved, and finally efficient and safe cooperative automatic driving is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and particularly relates to a vehicle-road-cloud dynamic collaborative autonomous driving method based on a space-ground integrated communication network. Background Art

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

[0003] Currently, there are still many challenges in autonomous driving technology in terms of wide-area coverage, real-time data transmission, and collaborative control. Existing autonomous driving systems mainly rely on ground infrastructure and limited communication networks to achieve information exchange and collaborative decision-making among vehicles, roads, and clouds. However, these systems have certain limitations in terms of wide-area coverage, signal stability, and data transmission delay. The coverage range of ground communication base stations is limited, making it difficult to ensure stable communication in remote or complex terrain areas. At the same time, with the rapid increase in the number of autonomous vehicles, existing communication networks face problems of insufficient bandwidth and increased delay, affecting the effect of real-time collaborative control.

[0004] As an emerging communication technology, low-earth orbit (LEO) satellite communication has the advantages of wide-area coverage, low latency, and high reliability, and has gradually become an important means to supplement ground communication networks. In recent years, with the rapid deployment of LEO satellite constellations and the continuous progress of satellite communication technology, LEO satellites have shown great potential in providing efficient communication services globally. However, currently, LEO satellites have not fully utilized their advantages in the deep integration and collaborative work with ground systems. Especially in the field of autonomous driving, how to effectively integrate LEO satellites with ground communication networks to achieve wide-area collaboration of space-ground integration remains a technical problem to be solved urgently. Summary of the Invention

[0005] Aiming at the deficiencies in the prior art, the present invention provides a vehicle-road-cloud dynamic collaborative autonomous driving method based on a space-ground integrated communication network, aiming to build a more extensive coverage, more stable communication, and more efficient collaborative autonomous driving support system by integrating LEO satellites and ground infrastructure. This system can not only make up for the deficiencies of ground communication networks and improve the operation ability of autonomous vehicles in complex and wide-area environments, but also further improve the safety and reliability of autonomous driving systems through efficient data transmission and collaborative control.

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

[0007] Vehicle-Road-Cloud Dynamic Cooperative Autopilot Method Based on Space-Ground Integrated Communication Network:

[0008] The vehicle terminal starts, enables the network interface, and accesses the regional cloud through the communication dual-link; the communication dual-link is that the vehicle terminal simultaneously accesses the mobile communication network and the satellite communication network.

[0009] The regional cloud dynamically groups the vehicles and determines the vehicle-road-cloud cooperation strategy based on the ground facility conditions; the ground facility conditions refer to whether there is a roadside device terminal. If there is a roadside device terminal, it is set as a three-level vehicle-road-cloud cooperation strategy, and the roadside device terminal is the grouping decision center, and the regional cloud conducts macro decision-making guidance; if there is no roadside device terminal, it is set as a two-level vehicle-cloud cooperation strategy, and the regional cloud is the grouping decision center, and delay compensation is enabled during decision-making.

[0010] When determining the vehicle-road-cloud cooperation strategy, the regional cloud docks with the central cloud to achieve wide-area cloud-cloud cooperation, optimizes the decision-making results using low-earth orbit satellite data and historical experience, and improves the overall decision-making performance through the memory bank and the teacher-student model.

[0011] Furthermore, the communication dual-link is divided into a main link and a backup link; when the vehicle terminal enters the handover 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 attempt to connect to the new regional cloud, and the backup link continues to perform data transmission 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; after the vehicle terminal drives out of the handover buffer area between different regional clouds, the backup link stops the data transmission task, and only the main link performs data transmission.

[0012] Furthermore, the regional cloud dynamically groups the vehicles: when a vehicle joins or exits, the grouping mechanism is immediately started. When there is no change in the number of vehicles, the grouping mechanism is started regularly. In the grouping mechanism, the regional cloud clusters the regional location, roadside device information, historical experience, and the perception confidence of the vehicles to obtain each cooperation group.

[0013] Even further, the perception confidence of the vehicle is calculated by the regional cloud based on the vehicle-end configuration and location information for the newly joined vehicle; the perception confidence S of the vehicle perception = a * S base + b * S history + c * S env , where a, b, and c are calculation weights, S base is the basic perception score of the vehicle, S history is the experience scoring function, S env is the environmental complexity score; and S base = S camera*X + S lidar *Y + S radar *Z, S camera , S lidar and S radar are the calculation weights of the in-vehicle camera, lidar, and millimeter-wave radar respectively, and X, Y, and Z are the numbers of the in-vehicle camera, lidar, and millimeter-wave radar arranged on the vehicle respectively; M is the historical target detection accuracy of the vehicle; S env Obtaining method of: The regional cloud identifies the total number of vehicles N in the region according to satellite observations, combines with the road length L of this region, and calculates the traffic flow density accordingly and normalizes it, and sums the normalized traffic flow density and the weather level with weights to obtain the environmental complexity score.

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

[0015] The automotive terminal and the roadside device terminal detect the performance of V2X communication in the detection field and agree on a data fusion strategy. If the bandwidth is sufficient and the latency is low, then it is agreed to adopt the early fusion strategy. If the bandwidth is tight or the latency is high, then it is agreed to adopt the mid-term fusion strategy;

[0016] In the early fusion strategy, the automotive terminal shares its own original perception data, including lidar point cloud, camera image, and positioning information, with the roadside device terminal, and at the same time makes a single-vehicle decision based on its own perception data, that is, the next driving path; after receiving the data of other terminals in the group, the roadside device terminal performs spatio-temporal alignment on the lidar point cloud data according to the positioning information of other terminals and executes early fusion, encodes the fused data into the point cloud BEV feature A, maps and encodes the camera image data into the visual BEV feature A; then, fuses the point cloud BEV feature A and the visual BEV feature A to form the fused BEV feature A, and uploads it to the regional cloud; the roadside device terminal performs target detection and prediction based on the fused BEV feature A to generate a preliminary decision result, that is, the next driving path of all automotive terminals in this group;

[0017] In the mid-term fusion strategy, the automotive terminal first extracts features from its own perception data and performs multi-modal feature fusion to form the single-vehicle BEV feature A. After further encoding to reduce the size, it is shared with the roadside device terminal, and at the same time makes a single-vehicle decision based on its own perception data, that is, the next driving path; after receiving the data of other terminals in the group, the roadside device terminal performs spatio-temporal alignment on the single-vehicle BEV feature A according to the positioning information of other terminals and executes mid-term fusion to form the fused BEV feature B, and uploads it to the regional cloud; the roadside device terminal performs target detection and prediction based on the fused BEV feature B to generate a preliminary decision result, that is, the next driving path of all automotive terminals in this group;

[0018] After receiving the fused BEV feature A or fused BEV feature B from the roadside device end within the packet, the regional cloud makes corrections by combining the observation data of the low-earth orbit satellite and historical experience; then, the regional cloud performs target detection and prediction to generate a regional decision result, and at the same time requests the macro traffic flow optimization target and decision guidance from the central cloud;

[0019] After receiving the request from the regional cloud, the central cloud generates a globally optimal traffic organization plan through large-scale traffic simulation and optimization algorithms, and then sends the macro traffic flow optimization target and decision guidance to the regional cloud; based on this, the regional cloud optimizes the regional decision result and sends it back to the roadside device end within the packet through the main link, so as to issue a timing or adjustment instruction to the current roadside device end;

[0020] The roadside device end optimizes its own preliminary decision result based on the regional decision optimization result and sends it to the vehicle terminal within the packet;

[0021] After receiving the packet decision result, the vehicle terminal optimizes its own single-vehicle decision result based on this, then outputs waypoints according to the final decision result, calculates the speed and direction based on the vehicle kinematic model and dynamic model, and finally outputs a control signal for vehicle control;

[0022] The sufficiency or tightness of the bandwidth therein is determined by the number of vehicle terminals and roadside device ends in the field; while the higher or lower latency is determined by the idle capacity of the bandwidth.

[0023] Furthermore, the correction by combining the observation data of the low-earth orbit satellite and historical experience is specifically as follows: the regional cloud retrieves the satellite observation data according to the geographical location information of the vehicle and the road, maps the satellite observation data to the grid map of the current road section through accurate spatio-temporal coordinate alignment. At the same time, the regional cloud stores the vehicle driving data, perception records and decision results in this area at different times and in different environments. The regional cloud compares the current BEV feature A or fused BEV feature B with similar scenarios in historical experience, calibrates the scene perception error therein, and comprehensively aligns and weights the fused BEV feature A or fused BEV feature B, satellite observation data and historical experience results, and outputs a corrected environment perception result as the input for subsequent target detection and prediction.

[0024] Furthermore, the roadside device end optimizes its preliminary decision result based on the regional decision optimization result, specifically as follows: First, an optimization objective is established and constraints are set. Then, a local optimization algorithm is adopted to perform secondary solution according to the macro guidance issued by the region. At the same time, local fine-tuning is performed on the speed suggestions, lane allocation, and queue sorting of grouped vehicles. If a vehicle has an emergency priority requirement, a higher weight is assigned to it during the local fine-tuning process. Finally, the optimized decision result is sorted into specific instructions or suggestions that each vehicle in the group can execute.

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

[0026] After receiving the grouping result and the collaboration strategy, the vehicle terminal starts to detect the communication performance of the main link. If the bandwidth is sufficient and the latency is low, it negotiates with the regional cloud to adopt early-stage fusion. If the bandwidth is tight or the latency is high, it negotiates with the regional cloud to adopt mid-stage fusion;

[0027] In the early-stage fusion strategy, the vehicle terminal transmits its original perception data, including lidar point cloud, camera images, and positioning information, to the regional cloud through the main link, and at the same time makes a single-vehicle decision based on its own perception data; after receiving the data of each vehicle terminal, the regional cloud performs spatio-temporal alignment on the lidar point cloud data according to the positioning information of each vehicle terminal and executes early-stage fusion, encodes the fused data into the point cloud BEV feature B, maps and encodes the camera image data into the visual BEV feature B, and uses a multi-modal feature fusion algorithm to fuse the two to form the fused BEV feature C;

[0028] In the mid-stage fusion strategy, the vehicle terminal first extracts features and performs multi-modal feature fusion on its own perception data to form a single-vehicle BEV feature B. After further encoding to reduce the size, it is transmitted to the regional cloud through the main link, and at the same time makes a single-vehicle decision based on its own perception data; after receiving the data of each vehicle terminal, the regional cloud performs spatio-temporal alignment on the single-vehicle BEV feature B according to its positioning information and executes mid-stage fusion to form the fused BEV feature D;

[0029] The regional cloud further corrects the fused BEV feature C or the fused BEV feature D by combining the observation data of the low-earth orbit satellite and historical experience; then, the region performs target detection and prediction to generate a regional decision result, and at the same time requests the central cloud for the macro traffic flow optimization objective and decision guidance;

[0030] After receiving the request from the regional cloud, the central cloud combines a wider traffic flow model and global environmental information to generate macroscopic traffic flow optimization objectives and more specific decision-making guidance, and sends the macroscopic traffic flow optimization objectives and decision-making guidance to the regional cloud. Based on this, the regional cloud optimizes the regional decision-making results, maps and connects the received macroscopic objectives with the regional traffic status, road resources, and real-time traffic flow to ensure that local execution is consistent with the global objectives, and introduces a delay compensation module for further advanced optimization. Finally, according to the driving requirements within each group, the regional cloud generates group decision-making results and sends them back to the vehicle terminals within the corresponding group through the main link;

[0031] After receiving the group decision-making results, the vehicle terminal optimizes its own single-vehicle decision-making results based on this, then outputs waypoints according to the final decision-making results, calculates the speed and direction based on the vehicle kinematic model and dynamic model, and finally outputs a control signal for vehicle control.

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

[0033] The regional cloud connects to the network main road through an optical fiber link and accesses the central cloud;

[0034] The regional cloud uploads the generated regional decision-making results to the central cloud. At the same time, it observes the movements of each group within the region through low-earth orbit satellites, collects the movement trajectories of vehicle terminals and their impacts on the surrounding traffic environment, and obtains a decision-making score through evaluation from three aspects: safety, regional traffic efficiency, and timeliness. The decision-making score S decision =α*S safety +β*S efficiency +γ*S time , where α, β, and γ are calculation weights, S safety is the safety score, S efficiency is the regional traffic efficiency score, and S time is the timeliness score;

[0035] The regional cloud encodes the regional decision-making results and their decision-making 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 a sparsification operation at regular intervals. Through data reconstruction and model learning, it obtains the optimal or sub-optimal decision-making operations for each region and each scenario within it, and based on this, establishes a high-performance teacher large model and continuously updates it through continuous learning;

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

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

[0039] Furthermore, the regional cloud sparsifies its memory bank at regular intervals, enhances the memory priority of optimal or sub-optimal decisions in various scenarios through data reconstruction and model learning, thereby updating its memory bank. When encountering similar scenarios, it preferentially refers to similar decisions in its 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 behaviors to make preliminary decisions. At the same time, it observes the regional decision results, records 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 reconstructions, if the new scenarios and new decisions can be proven effective, they will gradually obtain higher memory priorities and can be preferentially invoked when the same or similar scenarios reappear in the future.

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

[0041] (1) The present invention accesses the regional cloud through a dual-link of a mobile communication network and a satellite communication network, ensuring stable collaborative connections for vehicles in various environments and improving the reliability and fault tolerance of communication;

[0042] (2) In the present invention, the regional cloud dynamically groups vehicles according to the real-time configuration, location information, and perception confidence of the vehicles, and adopts a clustering strategy to ensure that vehicles within the collaborative group can achieve optimal or sub-optimal collaborative driving effects, improving the overall traffic efficiency and safety;

[0043] (3) The present invention supports two strategies: vehicle-road-cloud three-level collaboration and vehicle-cloud two-level collaboration, and can automatically switch according to the presence or absence of roadside collaborative devices, adapting to different infrastructure conditions and enhancing the adaptability and scalability of the system;

[0044] (4) Through the pre-fusion and mid-fusion strategies, the present invention optimizes the data transmission and fusion methods according to the network bandwidth and latency conditions, improves the utilization efficiency of perception data and the accuracy of decision-making, and further optimizes the decision results by combining low-earth orbit satellite observation data and historical experience;

[0045] (5) The present invention realizes wide-area cloud-cloud collaboration through the docking of the regional cloud and the central cloud, accumulates and optimizes decision-making experience in various scenarios through the construction of the memory bank, and uses a high-performance teacher large model and knowledge distillation technology to improve the performance of the regional cloud model, achieving fast and efficient decision optimization. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

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

[0049] The present invention will be described in detail below in conjunction with the embodiments shown in the drawings. However, these embodiments do not limit the present invention, and any structural, method, or functional transformation made by those of ordinary skill in the art based on these embodiments is included in the protection scope of the present invention.

[0050] The present invention proposes a vehicle-road-cloud dynamic collaborative autonomous driving method and system based on a space-ground integrated communication network, which combines low-earth orbit satellite communication and ground mobile communication to achieve stable collaboration between vehicles, roads, and clouds in a wide area. This method stably connects the vehicle terminal and the cloud through a dual communication link, formulates vehicle-road-cloud collaboration strategies according to ground facility conditions (presence or absence of roadside equipment), and supports vehicle-road-cloud three-level collaboration and vehicle-cloud two-level collaboration modes; at the same time, the regional cloud docks with the central cloud to achieve wide-area cloud-cloud collaboration, optimizes the decision-making results using low-earth orbit satellite data and historical experience, and improves the overall decision-making performance through a memory bank and teacher-student models, ultimately maximizing the efficiency of collaborative autonomous driving. Through the space-ground integrated communication and collaboration mechanism, this method significantly improves the capabilities of vehicle-road-cloud collaboration in terms of wide-area coverage, communication stability, and collaborative decision-making, enhances the safety and reliability of the system, and ultimately realizes efficient and safe collaborative autonomous driving, with broad application prospects.

[0051] As Figure 1 、 2 shown, the vehicle-road-cloud dynamic collaborative autonomous driving method of the present invention based on a space-ground integrated communication network includes the following steps:

[0052] S1. The vehicle terminal starts, turns on the network interface, and accesses the regional cloud through a dual link;

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

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

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

[0056] S5. The wide-area cloud-cloud collaboration and model construction process.

[0057] Further explanation, the step S1 is specifically:

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

[0059] S12. The vehicle terminal detects network links, selects the best link and simultaneously accesses the mobile communication network and the satellite communication network (i.e., dual communication links), and then accesses the nearby area cloud through the dual communication links.

[0060] S13. The vehicle terminal and the area cloud periodically detect the reachability and latency of the dual communication links, and automatically switch the link with the lowest latency as the main link for data transmission, and the other is the backup link to undertake communication tasks when the main link is interrupted, the latency is too high, or the area cloud switches.

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

[0062] S15. After the vehicle terminal drives out of the handover buffer area between different area clouds, the backup link stops the data transmission task, and only the main link performs data transmission.

[0063] Further explanation, the specific content of step S2 is as follows:

[0064] S21. The area cloud continuously provides communication docking for vehicles in the area to realize the dynamic joining and exiting of vehicles.

[0065] S22. The area cloud conducts a preliminary evaluation of the newly joined vehicle according to the vehicle-end configuration and location information, and calculates the vehicle's perception confidence based on historical experience.

[0066] Suppose a newly joined vehicle A enters the coverage area of the area cloud. The area cloud first obtains the vehicle's hardware configuration, assigns calculation weights according to the types of in-vehicle sensors and obtains the basic perception score. Suppose the calculation weights of the on-vehicle camera, lidar, and millimeter-wave radar are S camera 、S lidar and S radar , and the numbers of the on-vehicle camera, lidar, and millimeter-wave radar 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 regional cloud queries the historical perception performance of the vehicle and obtains an empirical scoring function. Assuming that the historical target detection accuracy of vehicle A is M and the error rate is 1 - M, the empirical scoring function of vehicle A is The regional cloud identifies the total number of vehicles N in the region according to satellite observations, and obtains the road length L of the region, and calculates the traffic flow density based on this And perform normalization. At the same time, combine meteorological data to obtain the current regional weather conditions, and divide them into several levels (0 - 10). The higher the weather level W, the greater the perception disturbance. Then, the normalized traffic flow density and the weather level are weighted and summed to obtain the environmental complexity score S env , which is used to measure the perception difficulty of the environment where vehicle A is located; Finally, the regional cloud is based on the basic perception score S of vehicle A base , empirical scoring function S history And environmental complexity score S env To obtain the perception confidence S of the vehicle perception = a * S base + b * S history + c * S env , where a, b, and c are calculation weights;

[0067] S23. The regional cloud performs dynamic grouping on vehicles. When a vehicle joins or exits, the grouping mechanism is immediately started. When there is no change in the number of vehicles, the grouping mechanism is started regularly to achieve the stability of grouping and the timeliness of dynamic changes;

[0068] S24. In the grouping mechanism, the regional cloud clusters the regional location, roadside device information, historical experience, and the perception confidence of the vehicle to obtain each collaboration group, so as to ensure that each collaboration group can achieve the optimal or sub - optimal collaborative driving effect;

[0069] S25. The regional cloud sets the vehicle - road - cloud collaboration strategy according to the roadside device information within the scope of each collaboration group. If there are roadside collaborative devices, it is set as three - level vehicle - road - cloud collaboration, and the roadside end is the grouping decision center, and the regional cloud conducts macro - decision guidance; If the roadside collaborative device is missing, it is set as two - level vehicle - cloud collaboration, and the regional cloud is the grouping decision center, and delay compensation is enabled during decision - making;

[0070] S26. The regional cloud sends the grouping result and the collaboration strategy to each vehicle terminal and roadside device terminal, and starts to receive grouping confirmation information. After receiving the confirmation from the terminal, it enables the collaboration service for the confirmed terminal and continues to wait for the unconfirmed terminal. If the waiting time exceeds the threshold, the unconfirmed terminal is temporarily removed from the grouping and regrouped when the next round of grouping mechanism is started.

[0071] Further explanation, the specific content of step S3 is:

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

[0073] S32. After receiving the grouping result and the collaboration strategy, the vehicle terminals and roadside equipment terminals directly connect to other terminals in the group through field V2X communication;

[0074] S33. The vehicle terminals and roadside equipment terminals detect the performance of field V2X communication and agree on a data fusion strategy. If the bandwidth is sufficient and the latency is low, they agree to adopt the early fusion strategy. If the bandwidth is tight or the latency is high, they agree to adopt the mid-term fusion strategy; the sufficiency or tightness of the bandwidth is determined by the number of vehicle terminals and roadside equipment terminals in the field. The more vehicle terminals and roadside equipment terminals there are in the field, the tighter the bandwidth, and vice versa. The specific division threshold is obtained from the historical experience of the scenario; the higher or lower 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 division threshold is obtained from the historical experience of the scenario;

[0075] S34. In the early fusion strategy, the vehicle terminal directly shares its own original perception data, including lidar point cloud, camera image, and positioning information, with the roadside equipment terminal through field V2X communication, and at the same time makes a single-vehicle decision based on its own perception data, that is, the next driving path; after receiving the data of other terminals in the group, the roadside equipment terminal performs spatio-temporal alignment on the lidar point cloud data according to the positioning information of other terminals and performs early fusion, encodes the fused data into point cloud BEV feature A, and maps and encodes the camera image data into visual BEV feature A; uses a multi-modal feature fusion algorithm to fuse point cloud BEV feature A and visual BEV feature A to form a fused BEV feature A, and uploads it to the regional cloud; at the same time, the roadside equipment terminal performs target detection and prediction based on the fused BEV feature A to generate a preliminary decision result, that is, the next driving path of all vehicle terminals in the group;

[0076] S35. In the mid-term fusion strategy, the vehicle terminal first extracts features from its own perception data and performs multi-modal feature fusion to form a single-vehicle BEV feature A. After further encoding to reduce its size, it is shared with the roadside device through field V2X communication. At the same time, based on its own perception data, the vehicle terminal makes a single-vehicle decision, that is, the next driving path. After receiving the data of other terminals in the group, the roadside device performs spatio-temporal alignment on the single-vehicle BEV feature A according to the positioning information of other terminals and executes mid-term fusion (for example, using maxout for feature fusion, retaining the larger value at the corresponding position in multiple feature maps) to form a fused BEV feature B, and uploads it to the regional cloud. At the same time, the roadside device performs object detection and prediction based on the fused BEV feature B to generate a preliminary decision result, that is, the next driving path of all vehicle terminals in the group.

[0077] S36. After receiving the fused BEV features A / B from the roadside device in the group, the regional cloud further corrects them by combining the observation data of low-earth orbit satellites and historical experience. Then, the regional cloud performs object detection and prediction to generate a regional decision result, and at the same time requests macroscopic traffic flow optimization objectives (such as large-scale traffic guidance strategies across urban areas and road networks) and decision-making guidance (such as suggesting a certain traffic flow to divert from a specific exit, suggesting temporarily changing the speed limit of some sections, or even enabling the emergency lane) from the central cloud.

[0078] Among them, the correction by combining the observation data of low-earth orbit satellites and historical experience is specifically as follows: The regional cloud retrieves satellite observation data according to the geographical location information of vehicles and roads, and maps the satellite observation data to the grid map of the current road section (such as the BEV coordinate system) through precise spatio-temporal coordinate alignment. At the same time, the regional cloud stores the vehicle driving data, perception records, and decision results in this region at different times and different environments (weather, time period, traffic flow). The regional cloud compares the current BEV features A / B with similar scenarios (similar weather, road structure, traffic flow level, etc.) in historical experience, calibrates the scene perception error, and comprehensively aligns and weights the fused BEV features A / B, satellite observation data, and historical experience results, and outputs a corrected environmental perception result as the input for subsequent object detection and prediction, providing a more accurate environmental state description for the next regional decision.

[0079] S37. After receiving the request from the regional cloud, the central cloud makes comprehensive decisions based on traffic data and models with a wider scope (which can schedule multiple regional clouds or national-level traffic center information simultaneously). Through large-scale traffic simulation and optimization algorithms, it generates a globally optimal or approximately optimal traffic organization plan (such as vehicle diversion guidance, global signal linkage of traffic lights, emergency accident handling plans, etc.). Then, it sends the macroscopic 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 device terminals within the group through the main links, so as to issue timing or adjustment instructions to the current roadside device terminals, such as coordinating signal linkages with adjacent intersections or adjacent sections of the road, instructing vehicles to execute diversions or maintain vehicle speeds, etc.;

[0080] The regional cloud optimizes the regional decision results based on this. Specifically: it integrates and adjusts the macroscopic traffic optimization strategy sent by the central cloud with the existing local decision-making scheme of the regional cloud to ensure the consistency between the local strategy and the global objective;

[0081] S38. The roadside device terminals optimize their preliminary decision results based on the regional decision optimization results and send them to the vehicle terminals within the group through field V2X communication;

[0082] The roadside device terminals optimize their preliminary decision results based on the regional decision optimization results. Specifically: first, establish optimization objectives. Without violating the regional decision instructions, maximize local traffic efficiency and safety, reduce vehicle interactions and conflicts, avoid collision risks, improve the passing rate as much as possible, relieve congestion at bottlenecks, and take into account the priority strategies for special vehicles (such as ambulances, buses, etc.). Then, set constraint conditions, such as safety regulations and environmental restrictions like road speed limits, lane capacities, signal timings, road construction areas or no-go lines. After that, adopt local optimization algorithms (such as linear / nonlinear programming, graph search algorithms, reinforcement learning local decision-making modules, etc.), perform secondary solutions according to the macroscopic guidance issued by the region, and at the same time comprehensively consider factors such as safety, efficiency, and comfort, and make local fine-tuning of the speed suggestions, lane assignments, queue sorting, etc. for the grouped vehicles (that is, speed suggestions, lane assignments, queue sorting, etc. are used as input parameters for the local optimization algorithm). If a vehicle has an urgent priority requirement, give it a higher weight during the local fine-tuning process. Finally, organize the optimized decision results into specific instructions or suggestions that each vehicle within the group can execute (such as target lanes / paths, expected speed ranges, driving priorities or following queue information, etc.);

[0083] S39. After receiving the grouped decision results, the vehicle terminals optimize their single-vehicle decision results based on this. Then, according to the final decision results, they output waypoints, calculate speeds and directions based on the vehicle kinematic model and dynamic model, and finally output control signals to control the vehicles;

[0084] After receiving the grouped decision result, the vehicle terminal optimizes its own single-vehicle decision result based on this. Specifically, algorithms such as local reinforcement learning are used to comprehensively weigh the grouped decision result and the vehicle's local perception result, and generate the final decision result. If a sudden situation is detected locally (such as an obstacle invading the vehicle's lane or an emergency), the vehicle terminal can make corresponding adjustments to the grouped decision on the premise of following the safety-first principle.

[0085] Further explanation, the specific steps of step S4 are as follows:

[0086] S41. Under the vehicle-cloud two-level collaborative strategy, the regional cloud sends the grouped result and the collaborative strategy to the vehicle terminals in the group, and synchronously uploads them to the central cloud;

[0087] S42. After receiving the grouped result and the collaborative strategy, the vehicle terminal starts to detect the communication performance of the main link. If the bandwidth is sufficient and the latency is low, it negotiates with the regional cloud to adopt the early fusion strategy. If the bandwidth is tight or the latency is high, it negotiates with the regional cloud to adopt the mid-term fusion strategy;

[0088] S43. In the early fusion strategy, the vehicle terminal transmits its own original perception data, including lidar point cloud, camera images, and positioning information, to the regional cloud through the main link, and at the same time makes a single-vehicle decision based on its own perception data. After receiving the data of each vehicle terminal, the regional cloud performs spatio-temporal alignment on the lidar point cloud data according to the positioning information of each vehicle terminal and executes early fusion, encodes the fused data into the point cloud BEV feature B, maps and encodes the camera image data into the visual BEV feature B, and uses a multi-modal feature fusion algorithm to fuse the two to form the fused BEV feature C;

[0089] S44. In the mid-term fusion strategy, the vehicle terminal first extracts features and performs multi-modal feature fusion on its own perception data to form a single-vehicle BEV feature B. After further encoding to reduce the size, it is transmitted to the regional cloud through the main link, and at the same time makes a single-vehicle decision based on its own perception data; after receiving the data of each vehicle terminal, the regional cloud performs spatio-temporal alignment on the single-vehicle BEV feature B according to its positioning information and executes mid-term fusion to form the fused BEV feature D;

[0090] S45. The regional cloud further corrects the fused BEV feature C / D by combining the observation data of the low-earth orbit satellite and historical experience, then performs object detection and prediction to generate the regional decision result, and at the same time requests the central cloud for the macroscopic traffic flow optimization goal and decision guidance;

[0091] S46. After receiving the requests from the regional cloud, the central cloud combines a broader traffic flow model and global environmental information to generate macroscopic traffic flow optimization objectives (e.g., diversion strategies, main artery priority passing plans, long-distance vehicle fleet scheduling rules, etc.) and more specific decision-making guidance (such as speed limits on specific sections of the road, preferential guidance for specific vehicle groups to pass, restricting the traffic flow in certain directions, etc.), and sends the macroscopic traffic flow optimization objectives and decision-making guidance to the regional cloud. The regional cloud optimizes the regional decision results based on this (mapping and connecting the received macroscopic objectives with the regional traffic status, road resources, real-time traffic flow, etc. information to ensure that local execution is consistent with the global objectives), and introduces the existing delay compensation module for further advanced optimization. Finally, it generates grouped decision results according to the driving requirements within each group and sends them back to the vehicle terminals within the corresponding group through the main link;

[0092] Generate grouped decision results according to the driving requirements within each group. Specifically, the regional cloud combines the differences in the destinations, current paths, travel attributes (private vehicles, commercial vehicles, public transportation, etc.) and time requirements (such as the expected arrival time) of the grouped vehicles to determine which vehicle groups need to have priority for smooth passage, which vehicle groups can be diverted or wait temporarily. At the same time, it uses algorithms such as linear / nonlinear programming, reinforcement learning, and model predictive control for comprehensive solution to generate executable specific instructions or suggestions for each vehicle (such as the target lane / path, expected speed range, driving priority or following queue information, etc.);

[0093] S47. After receiving the grouped decision results, the vehicle terminal optimizes its own single-vehicle decision results based on this (similar to S39), generates the final decision results, outputs waypoints, and then calculates the speed and direction based on the vehicle kinematic model and dynamic model, and finally outputs control signals to control the vehicle.

[0094] For further explanation, see Figure 3 , the specific steps of S5 are as follows:

[0095] S51. The regional cloud connects to the network main artery through an optical fiber link, accesses the central cloud, and realizes the interaction of data inside and outside the cloud through the central cloud to further improve its own 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 movements of each group within the region through low-earth orbit satellites, collects the movement trajectories of vehicle terminals and their impacts on the surrounding traffic environment, and obtains a decision score through evaluation from three aspects: safety, regional traffic efficiency, and timeliness;

[0097] Decision scores are obtained through evaluation in terms of three aspects: safety, regional traffic efficiency, and timeliness. Specifically: First, based on the collected collision risk index, the frequency of rapid acceleration / rapid braking, and the number of warnings for insufficient minimum vehicle distance, they are indexed, weighted, and averaged, and then normalized to obtain the safety score S safety ; Taking the average vehicle speed, vehicle passing rate, and intersection queue duration / length as the core indicators, they are normalized by comparing with historical averages or expected thresholds to obtain the regional traffic efficiency score S efficiency ; Index and weight the comprehensive decision-making delay, command execution delay, and traffic state improvement delay, and average them to obtain the timeliness score S time ; Finally, the indicators in the three dimensions are weighted and integrated to obtain the decision score S decision =α*S safety +β*S efficiency +γ*S time , where α, β, and γ are calculation weights;

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

[0099] S54. The regional cloud periodically performs a sparsification operation on its own memory bank, enhances the memory priority of the optimal or sub-optimal decisions in each scenario through data reconstruction and model learning, thereby updating its own memory bank. When encountering a similar scenario, it can preferentially refer to similar decisions in its own memory bank to improve the performance and reaction speed of decision-making; 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 behaviors, etc., so that the system can make a preliminary decision when there is no available ready-made experience; at the same time, observe the decision result, record the decision-making process and its execution result in its own memory bank to record and accumulate new experience. In subsequent sparsification operations or data reconstructions, if these new scenarios and new decisions can be proven effective, they will gradually obtain a higher memory priority and can be preferentially called when the same or similar scenarios appear again 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 periodically performs a sparsification operation to obtain the optimal or sub-optimal decision-making operations in each region and each scenario within it through data reconstruction and model learning, based on which a high-performance teacher large model is established and continuously updated through continuous learning;

[0101] S56. After establishing the high-performance teacher large model, the central cloud performs knowledge distillation on each regional cloud to guide the training of the student small models of each regional cloud and improve the performance of the regional cloud models;

[0102] S57. The regional cloud quickly makes optimal or sub-optimal decisions based on the student small models and its own memory bank, and transmits them back to each group through the main link, ultimately achieving efficient and secure wide-area vehicle-road-cloud collaborative autonomous driving.

[0103] The series of detailed descriptions listed above are only specific descriptions of the feasible implementation modes of the present invention, and they are not intended to limit the protection scope of the present invention. Any equivalent implementation modes or changes made without departing from the technical spirit of the present invention should be included within the protection scope of the present invention.

Claims

1. A vehicle-road-cloud dynamic collaborative autonomous driving method based on a space-ground integrated communication network, characterized in that: The vehicle terminal starts, enables the network interface, and accesses the regional cloud through a communication dual-link; the communication dual-link is that the vehicle terminal simultaneously accesses the mobile communication network and the satellite communication network; The regional cloud dynamically groups the vehicles and determines the vehicle-road-cloud collaboration strategy based on the ground facility conditions; the ground facility conditions refer to whether there is a roadside device terminal. If there is a roadside device terminal, it is set as a three-level vehicle-road-cloud collaboration strategy, and the roadside device terminal is the grouping decision center, and the regional cloud conducts macro decision-making guidance; if there is no roadside device terminal, it is set as a two-level vehicle-cloud collaboration strategy, and the regional cloud is the grouping decision center, and delay compensation is enabled during decision-making; When determining the vehicle-road-cloud collaboration strategy, the regional cloud docks with the central cloud to achieve wide-area cloud-cloud collaboration, uses low-orbit satellite data and historical experience to optimize the decision result, and improves the overall decision-making performance through the memory bank and the teacher-student model.

2. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 1, wherein, The communication dual-link is divided into a main link and a backup link; when the vehicle terminal enters the handover 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 attempt to connect to the new regional cloud, and the backup link continues to perform data transmission 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; after the vehicle terminal drives out of the handover buffer area between different regional clouds, the backup link stops the data transmission task, and only the main link performs data transmission.

3. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 1, wherein, The regional cloud dynamically groups the vehicles: when a vehicle joins or exits, the grouping mechanism is immediately started. When there is no change in the number of vehicles, the grouping mechanism is started regularly. In the grouping mechanism, the regional cloud clusters the regional location, roadside device information, historical experience, and the perception confidence of the vehicle to obtain each collaboration group.

4. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 3, wherein The perception confidence of the vehicle is calculated by the regional cloud for the newly added vehicle based on the vehicle-end configuration and location information; the perception confidence S of the vehicle perception = a * S base + b * S history + c * S env , where a, b, and c are calculation weights, S base is the basic perception score of the vehicle, S history is the experience scoring function, S env is the environmental complexity score; and S base = S camera * X + S lidar * Y + S radar * Z, S camera , S lidar and S radar are the calculation weights of the on-vehicle camera, lidar, and millimeter-wave radar respectively, and X, Y, and Z are the numbers of the on-vehicle camera, lidar, and millimeter-wave radar arranged on the vehicle respectively; M is the historical target detection accuracy of the vehicle; The obtaining method of S env : The regional cloud identifies the total number N of vehicles in the region according to satellite observations, combines with the road length L of the region, and calculates the traffic flow density and normalizes it, and sums the normalized traffic flow density and the weather level with weights to obtain the environmental complexity score.

5. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 1, wherein The three-level vehicle-road-cloud collaboration strategy is as follows: The vehicle terminal and the roadside device terminal detect the performance of the field V2X communication and agree on a data fusion strategy. If the bandwidth is sufficient and the latency is low, it is agreed to adopt the early fusion strategy. If the bandwidth is tight or the latency is high, it is agreed to adopt the mid-term fusion strategy; In the early fusion strategy, the vehicle terminal shares its own original perception data, including lidar point cloud, camera image, and positioning information, with the roadside device terminal, and at the same time makes a single-vehicle decision based on its own perception data, that is, the next driving path; after the roadside device terminal receives the data of other terminals in the group, it performs spatio-temporal alignment on the lidar point cloud data according to the positioning information of other terminals and performs early fusion, encodes the fused data into point cloud BEV feature A, and maps and encodes the camera image data into visual BEV feature A; then, fuses the point cloud BEV feature A and the visual BEV feature A to form a fused BEV feature A and uploads it to the regional cloud; the roadside device terminal performs target detection and prediction based on the fused BEV feature A to generate a preliminary decision result, that is, the next driving path of all vehicle terminals in the group; In the mid-term fusion strategy, the vehicle terminal first extracts features from its own perception data and performs multi-modal feature fusion to form a single-vehicle BEV feature A. After further encoding to reduce its size, it is shared with the roadside device side. At the same time, based on its own perception data, the vehicle terminal makes a single-vehicle decision, that is, the next driving path; after receiving the data of other terminals in the group, the roadside device side performs spatio-temporal alignment on the single-vehicle BEV feature A according to the positioning information of other terminals and executes mid-term fusion to form a fused BEV feature B, which is uploaded to the regional cloud; the roadside device side performs object detection and prediction based on the fused BEV feature B to generate a preliminary decision result, that is, the next driving path of all vehicle terminals in the group. After receiving the fused BEV feature A or fused BEV feature B of the roadside device side in the group, the regional cloud corrects it by combining the observation data of the low-earth orbit satellite and historical experience; then, the regional cloud performs object detection and prediction to generate a regional decision result, and at the same time requests the macro traffic flow optimization target and decision guidance from the central cloud. After receiving the request from the regional cloud, the central cloud generates a globally optimal traffic organization plan through large-scale traffic simulation and optimization algorithms, and then sends the macro traffic flow optimization target and decision guidance to the regional cloud; based on this, the regional cloud optimizes the regional decision result and sends it back to the roadside device side in the group through the main link, so as to issue a timing or adjustment instruction to the current roadside device side. The roadside device side optimizes its own preliminary decision result based on the regional decision optimization result and sends it to the vehicle terminals in the group. After receiving the group decision result, the vehicle terminal optimizes its own single-vehicle decision result based on this, then outputs waypoints according to the final decision result, calculates the speed and direction based on the vehicle kinematic model and dynamic model, and finally outputs a control signal to control the vehicle. The sufficiency or tightness of the bandwidth therein is determined by the number of vehicle terminals and roadside device sides in the field; while the high or low latency is determined by the idle capacity of the bandwidth.

6. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 5, wherein, The correction by combining the observation data of the low-earth orbit satellite and historical experience is specifically as follows: the regional cloud retrieves the satellite observation data according to the geographical location information of the vehicle and the road, maps the satellite observation data to the grid map of the current road section through accurate spatio-temporal coordinate alignment. At the same time, the regional cloud stores the vehicle driving data, perception records and decision results in this region at different times and different environments. The regional cloud compares the current BEV feature A or fused BEV feature B with similar scenarios in historical experience, calibrates the scene perception error therein, and comprehensively aligns and weights the fused BEV feature A or fused BEV feature B, satellite observation data and historical experience results, and outputs a corrected environmental perception result as the input for subsequent object detection and prediction.

7. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 6, wherein, The roadside device end optimizes its preliminary decision result based on the regional decision optimization result. Specifically, it first establishes an optimization goal and sets constraint conditions, and then adopts a local optimization algorithm to perform secondary solution according to the macro guidance issued by the region. At the same time, it makes local fine-tuning of the speed suggestions, lane assignments, and queue sorting of grouped vehicles. If a vehicle has an emergency priority requirement, a higher weight is assigned to it during the local fine-tuning process. Finally, the optimized decision result is sorted into specific instructions or suggestions that can be executed by each vehicle within the group.

8. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 7, wherein The vehicle-cloud two-level collaboration strategy is as follows: After receiving the grouping result and the collaboration strategy, the vehicle terminal starts to detect the communication performance of the main link. If the bandwidth is sufficient and the latency is low, it agrees with the regional cloud to adopt early fusion. If the bandwidth is tight or the latency is high, it agrees with the regional cloud to adopt mid-term fusion; In the early fusion strategy, the vehicle terminal transmits its original perception data, including lidar point cloud, camera images, and positioning information, to the regional cloud through the main link, and at the same time makes single-vehicle decisions based on its own perception data. After receiving the data of each vehicle terminal, the regional cloud performs spatio-temporal alignment on the lidar point cloud data according to the positioning information of each vehicle terminal and executes early fusion. The fused data is encoded into the point cloud BEV feature B, the camera image data is mapped and encoded into the visual BEV feature B, and the two are fused using a multi-modal feature fusion algorithm to form the fused BEV feature C; In the mid-term fusion strategy, the vehicle terminal first extracts features and performs multi-modal feature fusion on its own perception data to form the single-vehicle BEV feature B. After further encoding and reducing the size, it is transmitted to the regional cloud through the main link, and at the same time makes single-vehicle decisions based on its own perception data. After receiving the data of each vehicle terminal, the regional cloud performs spatio-temporal alignment on the single-vehicle BEV feature B according to its positioning information and executes mid-term fusion to form the fused BEV feature D; The regional cloud further corrects the fused BEV feature C or the fused BEV feature D by combining the observation data of the low-earth orbit satellite and historical experience; then, the region performs target detection and prediction to generate a regional decision result, and at the same time requests the macro traffic flow optimization goal and decision guidance from the central cloud; After receiving the request from the regional cloud, the central cloud combines a wider range of traffic flow models and global environment information to generate a macro traffic flow optimization goal and more specific decision guidance, and sends the macro traffic flow optimization goal and decision guidance to the regional cloud. Based on this, the regional cloud optimizes the regional decision result, maps and connects the received macro goal with the regional own traffic state, road resources, and real-time traffic volume to ensure that the local execution is consistent with the global goal, and introduces a delay compensation module to further optimize in advance. Finally, it generates a grouped decision result according to the driving requirements within each group and sends it back to the vehicle terminals within the corresponding group through the main link; After receiving the grouped decision result, the vehicle terminal optimizes its own single-vehicle decision result based on this, then outputs waypoints according to the final decision result, calculates the speed and direction based on the vehicle kinematic model and dynamic model, and finally outputs a control signal to control the vehicle.

9. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 1, wherein The regional cloud docking center cloud is used to achieve wide-area cloud-cloud collaboration, specifically as follows: The regional cloud is connected to the network main road through a fiber optic link and accesses the central cloud; The regional cloud uploads the generated regional decision results to the central cloud. At the same time, it observes the movements of each group within the region through low-earth orbit satellites, collects the movement trajectories of vehicle terminals and their impacts on the surrounding traffic environment, and obtains a decision score by evaluating from three aspects: safety, regional traffic efficiency, and timeliness; the decision score S decision = α * S safety + β * S efficiency + γ * S time , where α, β, and γ are calculation weights, S safety is the safety score, S efficiency is the regional traffic efficiency score, and S time is the timeliness score; The regional cloud encodes the regional decision results and their decision scores into memory features, stores them in its own memory bank, and simultaneously uploads them to the central cloud synchronously; After receiving the memory features of each regional cloud, the central cloud stores them in the global memory bank, performs a sparsification operation at regular intervals, obtains the optimal or sub-optimal decision operations for each region and each internal scenario through data reconstruction and model learning, establishes a high-performance teacher large model based on this, and continuously updates it through continuous learning; After establishing the high-performance teacher large model, the central cloud performs knowledge distillation on each regional cloud to guide the training of the student small models of each regional cloud; Based on the student small model and its own memory bank, the regional cloud quickly makes optimal or sub-optimal decisions and sends them back to each group through the main link, ultimately achieving efficient and safe wide-area vehicle-road-cloud collaborative autonomous driving.

10. The vehicle-road-cloud dynamic collaborative autonomous driving method according to claim 9, wherein, The regional cloud performs a sparsification operation on its own memory bank at regular intervals, enhances the memory priority of the optimal or sub-optimal decisions in each scenario through data reconstruction and model learning, thereby updating its own memory bank. When encountering a similar scenario, it preferentially refers to the similar decisions in its own memory bank. For new scenarios that have never been encountered or are not common, the regional cloud performs online learning or short-term reinforcement learning on the collected sensor data and vehicle dynamic behaviors to make a preliminary decision; at the same time, it observes the regional decision results, records the decision process and its execution results in the local memory bank to record and accumulate new decisions. In subsequent sparsification operations or data reconstructions, if the new scenario and new decision can be proven effective, they will gradually obtain a higher memory priority, and in the future, when the same or similar scenario appears again, they can be preferentially invoked.

Images