Multi-vehicle cooperative control method and system based on edge and center cloud combined computing

Abstract

The invention discloses a multi-vehicle cooperative control method and system based on edge and center cloud combined computing. The system comprises an unmanned driving module and an external module. An environment sensing module, a decision planning module, a vehicle-mounted communication module and a control execution module are arranged in the unmanned driving module; an edge computing cloud module and a central cloud computing module are arranged in the external module; the decision planning module is connected with the environment sensing module, the vehicle-mounted communication module and the control execution module. The edge computing cloud module is connected with the central cloud computing module; the edge computing cloud module and the central cloud computing module are respectively connected with the decision planning module and the environment sensing module through the vehicle-mounted communication module; and the environment sensing module is also connected with the control execution module. The real-time performance, the communication efficiency and the system stability of the multi-vehicle unmanned driving system can be improved, and meanwhile the cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of unmanned driving technology, and in particular to a multi-vehicle collaborative control method and system based on edge and central cloud joint computing. Background Art

[0002] Multi-vehicle autonomous driving is an important branch of the current autonomous driving technology field that has attracted much attention. It contains many technical challenges and potential social application value. With the continuous development of autonomous driving technology, multi-vehicle autonomous driving systems are leading the transformation of the transportation field. This technical field aims to achieve collaborative operation between multiple autonomous vehicles, aiming to improve the efficiency of road traffic, reduce traffic congestion, enhance road safety, and improve users' travel experience. However, multi-vehicle autonomous driving faces a series of complex technical challenges in practical applications.

[0003] First, communication and cooperative control are the core issues of multi-vehicle autonomous driving. Real-time communication and cooperative operation between multiple autonomous vehicles requires highly reliable communication networks and advanced cooperative control algorithms. These algorithms must ensure that vehicles can coordinate with each other, avoid collisions, and achieve efficient traffic flow. The challenge here is how to achieve fast and reliable data transmission to ensure that vehicles can respond quickly to changing traffic conditions.

[0004] Secondly, multi-vehicle autonomous driving requires processing a large amount of data. Each autonomous vehicle is equipped with a variety of sensors, such as cameras, lidar, GPS, etc., to perceive the surrounding roads and traffic conditions. The massive data generated by these sensors needs to be processed in real time for decision-making and path planning. Therefore, the computing task of processing this data is huge and requires powerful computing resources to support it.

[0005] Furthermore, decision-making and planning issues are also core challenges for multi-vehicle autonomous driving. In complex traffic environments, autonomous vehicles need to make intelligent decisions and plan paths, taking into account factors such as the behavior of other vehicles, traffic signals, and road conditions. This requires advanced decision-making algorithms and path planning technologies to ensure that vehicles can travel safely and efficiently.

[0006] Therefore, based on the above background, how to achieve coordinated control of multiple unmanned vehicles has become a technical problem that needs to be urgently solved by technical personnel in this field. Summary of the invention

[0007] The main purpose of the present invention is to provide a multi-vehicle collaborative control method and system based on edge and central cloud joint computing, aiming to improve the real-time performance, communication efficiency and system stability of the multi-vehicle unmanned driving system while reducing costs.

[0008] To achieve the above objectives, the present invention provides a multi-vehicle collaborative control method based on edge and central cloud joint computing.

[0009] The multi-vehicle cooperative control method based on edge and central cloud joint computing includes the following steps:

[0010] Step 1: When any unmanned vehicle A starts, its target destination position is given; the central cloud computing module is used to control the unmanned vehicle A and receive vehicle data; at the same time, the central cloud computing module sends the map and traffic status data to the edge computing cloud module a corresponding to the unmanned vehicle A, and the edge computing cloud module a sends the processed map and traffic status data to the unmanned vehicle A;

[0011] Step 2: Unmanned vehicle A uses a global path planning algorithm to integrate the map processed by edge computing cloud module a, traffic condition data, and the location data of the unmanned vehicle itself to generate a global path that will guide unmanned vehicle A from the starting point to the target location;

[0012] Step 3: The central cloud computing module calculates the correlation between unmanned vehicle A and other unmanned vehicles according to the request of unmanned vehicle A, and sends the environmental data perceived by the highly correlated unmanned vehicles to the edge computing cloud module a; at the same time, unmanned vehicle A sends its own perceived environmental data to the edge computing cloud module a; after receiving the environmental data, the edge computing cloud module a processes and updates it and sends it to the central cloud computing module;

[0013] Step 4: Unmanned vehicle A uses the processed environmental data, combined with its own linear speed and angular speed data, to perform local path planning, and sends the vehicle driving status data to the central cloud computing module; at the same time, when unmanned vehicle A drives to the area of ​​other edge computing cloud modules, the edge computing cloud modules interact with each other;

[0014] Step 5: Unmanned vehicle A uses MPC model-based predictive control to move along the reference point of the optimal trajectory;

[0015] Step 6: During the movement, determine whether there are no dynamic obstacles on the moving path. If so, return to step 3, reacquire the correlation information, update the global path and local path, and then re-perform motion control; if not, continue to determine whether it has moved to the end of the trajectory. If so, then determine whether it has reached the end. If it has not reached the end, return to step 3. If it has reached the end, output the path; otherwise, continue to predict control based on the MPC model and repeat step 5.

[0016] In one embodiment, the global path planning algorithm is an improved A-star algorithm, and the steps of performing global path planning by the improved A-star algorithm are as follows:

[0017] Step 2.1, unmanned vehicle A receives map and traffic condition data, and obtains its own location data;

[0018] Step 2.2, rasterize and expand the map, then initialize the open list and closed list in the A-star algorithm, and put the starting point into the open list;

[0019] Step 2.3, traverse the open list to find the node corresponding to the minimum evaluation function, put it into the closed list, and set it as the current node N;

[0020] Step 2.4: Determine whether the current node N is the end point. If so, output the planned path and jump to step 7; otherwise, if it is not the end point, proceed to step 5;

[0021] Step 2.5: Expand the current node N according to the corresponding expansion method, and put the expanded nodes that are not in the open list into the open list;

[0022] Step 2.6: Check whether the open list is empty. If it is empty, the path planning fails. Otherwise, if it is not empty, return to step 3.

[0023] Step 2.7: Use a second-order Bezier curve to smooth the planned path to obtain the optimal path.

[0024] In one embodiment, the correlation consists of a position distance correlation and a speed correlation;

[0025] The position distance correlation dis(A,B) is expressed as follows:

[0026]

[0027] Where: x_A and y_A represent the position coordinates of vehicle A, x_B and y_B represent the position coordinates of vehicle B;

[0028] The velocity correlation vel(A,B) is expressed as follows:

[0029] vel(A,B)=abs(v_A-v_B)

[0030] Where: v_A represents the speed of vehicle A, v_B represents the speed of vehicle B; abs() represents the absolute value function;

[0031] According to the position distance and speed correlation formula, the correlation formula cor(A,B) is defined as follows:

[0032] cor(A,B)=k1*dis(A,B)+k2*vel(A,B)

[0033] Among them, k1 and k2 are weight factors used to balance the influence of different factors.

[0034] In one embodiment, the step of acquiring the perceived environmental data is specifically as follows:

[0035] Step 3.1: Synchronize the multiple cameras and lidars of the unmanned vehicle in time and space to ensure that data is captured at the same timestamp;

[0036] Step 3.2: When the unmanned vehicle is moving, the camera collects images, the laser radar collects point cloud data, and then the images and point cloud data are preprocessed;

[0037] Step 3.3, performing image registration on the preprocessed image and performing image fusion; performing point cloud registration on the preprocessed point cloud data and performing fusion on multiple point cloud data;

[0038] Step 3.4, perform target detection on the fused image and point cloud data;

[0039] Step 3.5: Data association is performed between the object detection results of the image and the object detection results of the point cloud data to ensure that they correspond in time and space.

[0040] Step 3.6, performing target-level fusion on the associated target data;

[0041] Step 3.7: Output the fusion result, which includes the category, position, and speed of the target.

[0042] In one embodiment, the steps of performing local path planning are specifically as follows:

[0043] Step 4.1: Establish the motion model of the unmanned vehicle: The unmanned vehicle moves within the time interval Δt, and the kinematic model is:

[0044] x t+Δt =x t +v x Δtcosθ t -v y Δtsinθ t ;

[0045] y t+Δt =y t +v x Δtsinθ t +v y Δtcosθ t ;

[0046] θ t+Δt =θ t +ω t Δt;

[0047] In the formula, x t+Δt is the horizontal coordinate of the unmanned vehicle at time t+Δt, y t+Δt is the ordinate of the unmanned vehicle at time t+Δt, x t is the horizontal coordinate of the unmanned vehicle at time t, y t is the ordinate of the unmanned vehicle at time t, v x and v y is the linear velocity of the unmanned vehicle along the x and y axes at time t in the chassis coordinate system, θ t is the angle between the driverless car and the x-axis at time t, ω t is the angular velocity of the autonomous vehicle at time t;

[0048] Step 4.2, unmanned vehicle speed sampling: There are infinite sets of speed pairs (v, ω) in the two-dimensional speed space. According to the limitations of the unmanned vehicle itself and environmental factors, constraints are imposed to control the speed search space (v, ω) within a dynamic range, namely, the boundary limitation conditions:

[0049] v d ={(v,ω)|v∈[v min ,v max ]^ω∈[ω min ,ω max ]};

[0050] Where: v m is the maximum and minimum speed limits of the unmanned vehicle, v max and v min are the maximum and minimum linear speeds that the unmanned vehicle can achieve, ω max and ω min They are the maximum angular velocity and minimum angular velocity that the unmanned vehicle can reach respectively;

[0051] Due to the maximum acceleration and deceleration limit imposed by the driving force of the unmanned vehicle, the speed space is further constrained by v d Indicates; let the current linear velocity of the robot be v c , the current angular velocity is ω c , the time interval is Δt, then the achievable speed v at the next moment d for:

[0052]

[0053] In the formula, is the maximum acceleration; is the maximum deceleration; is the maximum angular acceleration; is the maximum angular deceleration;

[0054] In a time interval, the speed that the unmanned vehicle can reach is within a window, that is, the unmanned vehicle has a maximum acceleration limit:

[0055]

[0056] Where: dist(v,ω) represents the path length between the unmanned vehicle and the nearest obstacle;

[0057] Within a forward cycle of the unmanned vehicle, Δt, the effective speed set interval v is:

[0058] v=v m ∩v d ∩v a ;

[0059] Step 4.3: Use the evaluation function to evaluate the trajectories corresponding to the sampled groups of speeds to select the optimal trajectory; the criteria for trajectory evaluation are: accurately avoiding obstacles and approaching the target point in the shortest time; the evaluation function is:

[0060] G(v,ω)=α·head(v,ω)+β·dist(v,ω)+γ·vel(v,ω);

[0061] Where α, β and γ are the weight factors of each item, head(v,ω) is the azimuth evaluation function, which indicates the azimuth deviation between the direction of the unmanned vehicle and the end point of the entire path when moving to the end point of a planned path at the current sampling speed. The smaller the value, the higher the evaluation score; dist(v,ω) is the safety factor evaluation function, which indicates the path length between the unmanned vehicle and the nearest obstacle; vel(v,ω) is the current speed evaluation function.

[0062] In one embodiment, before performing speed evaluation, the azimuth evaluation function, the safety factor evaluation function and the current speed magnitude evaluation function are normalized, and the normalization process is as follows:

[0063]

[0064]

[0065]

[0066] Where n is the trajectory in the dynamic window, and i is the trajectory that needs to be evaluated.

[0067] In one embodiment, the steps of the MPC model predictive control are specifically as follows:

[0068] Step 5.1, prediction model: At the beginning of each control cycle, based on the current vehicle state and system model, predict the state and control input in a series of discrete time steps in the future;

[0069] Step 5.2, rolling optimization: use numerical optimization method to solve and obtain the optimal control input sequence;

[0070] Step 5.3, feedback correction: At the beginning of each new control cycle, measure the actual state of the vehicle and correct the initial state based on the feedback information.

[0071] In addition, to achieve the above-mentioned objectives, the present invention also provides a system for a multi-vehicle collaborative control method based on edge and central cloud joint computing, the system for a multi-vehicle collaborative control method based on edge and central cloud joint computing comprising an unmanned driving module and an external module; the unmanned driving module is provided with an environmental perception module, a decision-making and planning module, an on-board communication module and a control execution module; the external module is provided with an edge computing cloud module and a central cloud computing module; the decision-making and planning module is respectively connected to the environmental perception module, the on-board communication module and the control execution module; the edge computing cloud module is connected to the central cloud computing module; the edge computing cloud module and the central cloud computing module are respectively connected to the decision-making and planning module and the environmental perception module via the on-board communication module; the environmental perception module is also connected to the control execution module.

[0072] In one embodiment, the environment perception module includes multiple sensors, including cameras, laser radars, GPS and IMU sensors;

[0073] The decision-making and planning module is used to receive data sent from the environment perception and recognition module and environment and traffic condition data transmitted from the external module through the vehicle communication module, and perform global path planning and local path planning;

[0074] The vehicle communication module is used to realize high-speed data interaction between the unmanned vehicle and the edge computing cloud module and the central cloud computing module;

[0075] The control execution module is used to convert the decision and path planning results of the decision planning module into specific vehicle actions;

[0076] The edge computing cloud module is located outside the unmanned vehicle and is a distributed computing model that places computing and data storage at the edge of the network;

[0077] The central cloud computing module is equipped with a variety of cloud platforms, including a vehicle networking cloud platform, a transportation management cloud platform and a map cloud platform.

[0078] The beneficial effects that can be achieved by the present invention are:

[0079] 1. Reduce communication delay: The edge computing cloud module brings computing resources closer to the vehicle, reduces data transmission delays, improves quick response time, avoids collisions and optimizes traffic flow.

[0080] 2. Reduce network congestion: Large-scale Internet of Vehicles communications may cause network congestion, affecting the real-time transmission of data. The edge computing cloud module of the present invention can process and cache data near the vehicle, reduce network load, and ensure timely data transmission.

[0081] 3. Reduce the server burden of the central cloud computing module: The edge computing cloud module can process part of the data, reduce the server burden of the central cloud computing module, and reduce the cost and complexity of the server in the central cloud computing module.

[0082] 4. Support real-time updates: The edge computing cloud module can provide real-time data updates based on the actual situation around the vehicle, without waiting for batch updates from the server in the central cloud computing module, and can better adapt to traffic changes or handle emergencies in a timely manner.

[0083] 5. Save computing time: By associating data with other vehicles, data that is highly relevant to driverless vehicle A can be reused, thereby reducing the amount of calculation and processing time, thereby helping the driverless vehicle make planning decisions faster and ensure smooth and safe driving.

[0084] In summary, the data processing and decision-making mode based on edge and central cloud joint computing in the present invention not only improves the intelligence level and driving efficiency of unmanned vehicles, but also effectively reduces time waste and ensures that vehicles can travel more safely and efficiently in complex road environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 is a flowchart of a multi-vehicle cooperative control method according to an embodiment of the present invention;

[0086] Figure 2 is a flow chart of an improved A-star algorithm in an embodiment of the present invention;

[0087] Figure 3 is a schematic diagram of rasterizing and expanding a map in an embodiment of the present invention;

[0088] Figure 4 It is a schematic diagram of the use of the global path planning result of the improved A-star algorithm in the present invention;

[0089] Figure 5 It is a schematic diagram of the process of environmental perception of the present invention;

[0090] Figure 6 It is the YOLO target detection flow chart;

[0091] Figure 7 It is a flowchart of LiDAR Euclidean clustering;

[0092] Figure 8is a schematic diagram of a flow chart of local path planning in an embodiment of the present invention;

[0093] Fig. 9 This is the schematic diagram of MPC model predictive control;

[0094] Fig.10 is a schematic diagram of the application of the present invention in a road;

[0095] Fig.11 It is a structural principle diagram of the system of the present invention;

[0096] Fig.12 This is a schematic diagram before and after multi-sensor timestamp synchronization.

[0097] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0098] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0099] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

[0100] Embodiment 1: Multi-vehicle cooperative control method based on edge and central cloud joint computing, such as Figure 1 As shown, the following steps are included:

[0101] Step 1: When any unmanned vehicle A starts, its target destination position is given; the central cloud computing module is used to control the unmanned vehicle A and receive vehicle data; at the same time, the central cloud computing module sends the map and traffic status data to the edge computing cloud module a corresponding to the unmanned vehicle A, and the edge computing cloud module a sends the processed map and traffic status data to the unmanned vehicle A;

[0102] Step 2: Unmanned vehicle A uses a global path planning algorithm to integrate the map processed by edge computing cloud module a, traffic condition data and the location data of the unmanned vehicle itself to generate a global path that will guide unmanned vehicle A from the starting point to the target location. In this step, the global path planning algorithm is an improved A-star algorithm, such as Figure 2 As shown, the steps of performing global path planning using the improved A-star algorithm are as follows:

[0103] Step 2.1, unmanned vehicle A receives map and traffic condition data, and obtains its own location data;

[0104] Step 2.2: Rasterize and expand the map, then initialize the open list and closed list in the A-star algorithm, and put the starting point into the open list; in this step, in the grid map, the path generated by the mobile robot may intersect with the corners of obstacles or cross obstacles. Although the path planning seems successful, it does not fully consider the actual traffic capacity of the robot, so this so-called "successful" path is actually not feasible.

[0105] To solve these problems, the present invention proposes a method of introducing an expansion layer to handle obstacles in a grid map. By expanding the obstacles, an expansion area that exceeds the width of the robot is created at the edge of the obstacle, thereby effectively preventing the generated path from intersecting the obstacle, such as Figure 3 shown. Figure 3 (a) is a schematic diagram of the planned path intersecting or passing through obstacles. Figure 3 (b) is a schematic diagram of the improved map path planning. It can be found that adding an expansion layer of obstacles can effectively reduce the number of search nodes of the A* algorithm, that is, there is no need to search the nodes in the expansion layer. This can save the time cost of planning and improve the computational efficiency of the algorithm. At the same time, the introduction of the expansion layer can also ensure that the mobile robot will not collide with obstacles during the actual movement process, thereby ensuring the safety of the robot's movement.

[0106] Step 2.3, traverse the open list to find the node corresponding to the minimum evaluation function, put it into the closed list, and set it as the current node N;

[0107] Step 2.4: Determine whether the current node N is the end point. If so, output the planned path and jump to step 7; otherwise, if it is not the end point, proceed to step 5;

[0108] Step 2.5: Expand the current node N according to the corresponding expansion method, and put the expanded nodes that are not in the open list into the open list;

[0109] Step 2.6: Check whether the open list is empty. If it is empty, the path planning fails. Otherwise, if it is not empty, return to step 3.

[0110] Step 2.7: Use the second-order Bezier curve to smooth the planned path to obtain the optimal path. The calculation formula of the second-order Bezier curve in this step is as follows:

[0111] B(t)=(1-t) 2 P0+2t(1-t)P1+t 2 P2,t∈[0,1];”

[0112] Where P i is the position point, t is the time;

[0113] In this embodiment, the global path planning reference obtained by the A-star algorithm Figure 4 .

[0114] Step 3: The central cloud computing module calculates the correlation between unmanned vehicle A and other unmanned vehicles according to the request of unmanned vehicle A, and sends the environmental data perceived by the highly correlated unmanned vehicles to the edge computing cloud module a; at the same time, unmanned vehicle A sends its own perceived environmental data to the edge computing cloud module a; after receiving the environmental data, the edge computing cloud module a processes and updates it and sends it to the central cloud computing module;

[0115] Specifically, in this step, the central cloud computing module will process the data according to the data requests sent by the unmanned vehicle and the edge computing cloud module. The central cloud computing module sends the required data after processing to the edge computing cloud module, including real-time high-precision maps, road conditions, and environmental data and location information perceived by other vehicles. In addition, the central cloud computing module will also monitor the driving status of the unmanned vehicle, and perform route scheduling and remote command issuance as needed to ensure that the vehicle can drive safely in various situations.

[0116] The edge computing cloud module will respond to the data request sent by the unmanned vehicle and send the data processed by edge computing to the vehicle, including data from the central cloud computing module (such as road conditions, real-time high-precision maps, environmental data and location information perceived by other vehicles, etc.), as well as the results of processing the sensor data of the unmanned vehicle itself. These processes help improve the driving safety and efficiency of the unmanned vehicle. In addition, the edge computing cloud module will combine the data sent by the central cloud computing module and related perception results and other information to make real-time decisions and send the decision results to the unmanned vehicle. At the same time, the edge computing cloud module will also update high-precision maps, environmental data and other information, and synchronize the updated data to the central cloud computing module.

[0117] At the same time, the unmanned vehicle will send the environmental data sensed by its sensors to the edge computing cloud module, which will process and analyze the data in real time to provide more accurate environmental perception information and decision support. In addition, the unmanned vehicle will also send its driving status, location and other information to the central cloud computing for monitoring, route scheduling and other activities.

[0118] It should be noted that in this step, when the central cloud computing sends the data collected by other vehicles to the edge computing cloud module, it does not blindly receive the data of all vehicles. Generally, the data is sent selectively based on the correlation between the request and the requested vehicle, and its correlation consists of position distance correlation and speed correlation;

[0119] The position distance correlation dis(A,B) is expressed as follows:

[0120]

[0121] Where: x_A and y_A represent the position coordinates of vehicle A, x_B and y_B represent the position coordinates of vehicle B;

[0122] The velocity correlation vel(A,B) is expressed as follows:

[0123] vel(A,B)=abs(v_A-v_B)

[0124] Where: v_A represents the speed of vehicle A, v_B represents the speed of vehicle B; abs() represents the absolute value function;

[0125] According to the position distance and speed correlation formula, the correlation formula cor(A,B) is defined as follows:

[0126] cor(A,B)=k1*dis(A,B)+k2*vel(A,B)

[0127] Where k1 and k2 are weight factors used to balance the influence of different factors.

[0128] Only when the correlation value is higher than a certain threshold, the central cloud computing will send the environmental information, location and other data perceived by the vehicle. In addition, the unmanned vehicle will also send its own status and location information to the central cloud computing to provide a reference for the decision-making planning of other vehicles. The edge computing cloud will also process the environmental perception data sent by the unmanned vehicle, update the map and road conditions, and then synchronize it to the central cloud computing module to reduce the computing workload of its server and provide more real-time and accurate data for the unmanned vehicle.

[0129] Furthermore, in this step, if Figure 5 As shown, the perception of the environmental data includes the following steps:

[0130] Step 3.1: synchronize the multiple cameras and lidars of the unmanned vehicle in time and space to ensure that data is captured at the same timestamp; this includes camera calibration, lidar and camera joint calibration, as follows:

[0131] 1. Camera calibration

[0132] Camera calibration involves the conversion between multiple coordinate systems, including the world coordinate system ( w -X w Y w Z w ), camera coordinate system (O c -X c Y c Zc ), image coordinate system (o-xy) and pixel coordinate system (o o -uv). The world coordinate system is used to describe the absolute position of sensors such as cameras and target objects in space, with the origin O w The camera coordinate system is used to describe the three-dimensional coordinates of objects in space relative to the camera body. The origin O c is the camera optical center, X c and Y c is the axis parallel to the pixel plane, Z c The image coordinate system is used to describe the two-dimensional coordinates of objects in space in the camera projection plane during imaging. The origin is the intersection of the camera optical axis and the image plane. The unit is millimeters. The pixel coordinate system is used to describe the position of the pixel point on the camera imaging chip on the image. The origin is o o is the top left corner of the image.

[0133] The conversion relationship between the camera coordinate system and the image coordinate system is as follows:

[0134]

[0135] Among them, K c ∈R 3×3 , is the intrinsic parameter matrix of the camera, which is defined as:

[0136]

[0137] In the formula, f x 、f y are the focal lengths of the camera’s x-axis and y-axis, respectively; u0 and v0 are the optical centers of the image coordinate system’s u-axis and v-axis, respectively

[0138] Use radial distortion parameters to correct the coordinates:

[0139]

[0140] In the formula, x r and r is the horizontal and vertical coordinates after correcting the radial distortion; x b and b are the horizontal and vertical coordinates of the distortion; k1, k2, k3 are the radial distortion of the camera respectively; r is the distance from the point to the imaging center.

[0141] Use the tangential distortion parameters to correct the coordinates:

[0142]

[0143] In the formula, x t and tare the horizontal and vertical coordinates after correcting the tangential distortion; p1 and p2 are the tangential distortion of the camera respectively.

[0144] 2. Joint calibration of LiDAR and camera;

[0145] 2.1 Spatial synchronization;

[0146] The calibration of LiDAR and millimeter-wave radar is divided into internal parameter calibration and external parameter calibration. The internal parameter calibration mainly includes distance correction angle, rotation correction angle, vertical correction angle and horizontal offset factor. Generally, the internal hardware parameters of LiDAR and millimeter-wave radar are set by the manufacturer. The point cloud data collected by LiDAR can be directly used for mapping and target detection. Millimeter-wave radar can detect obstacles.

[0147] The purpose of camera extrinsic calibration and lidar extrinsic calibration is to establish the relationship between the sensor and the world coordinate system or other sensor coordinate systems through calibration. The calibration principle is to use the calibration object feature point association method to complete the joint calibration, that is, two heterogeneous sensors collect the same chessboard calibration plate, extract the coordinates of the feature points on the calibration plate in the respective sensor coordinate systems, and solve the joint calibration extrinsic matrix.

[0148] The point cloud coordinates of the target object surface in the laser radar coordinate system can be expressed as (X, Y, Z), and the pixel coordinates of the feature point in the pixel coordinate system can be expressed as (u, v), that is, the representation method of the three-dimensional coordinates on the image plane after removing the distance information, and the two are always represented as the same feature point. The calibration process is to record the sensor data, including timestamps, point cloud data, image data, etc., through the rosbag function package; output the recorded data to the calibration function package in the host computer, and then manually select the relationship between the three-dimensional space feature points and the corresponding points of the two-dimensional image in the calibration function package; finally, select a certain number of feature point pairs for iterative calculation to solve the rotation matrix and translation matrix.

[0149] The point cloud coordinates can be expressed as the conversion relationship between (X, Y, Z) and the pixel coordinate system, as shown below:

[0150]

[0151] In the formula, R l is the rotation matrix, T l is the translation matrix.

[0152] 2.2 Time Synchronization

[0153] Because all sensors run according to independent clock references, time synchronization of multiple sensors becomes one of the necessary conditions for multi-sensor information fusion. If the sampling frequency is not fixed or different, the multiple sampling data will be out of sync in time.

[0154] There are two main methods for multi-sensor time synchronization: hardware synchronization and software synchronization. Hardware synchronization means that multiple sensors trigger sampling at the same time; software synchronization means that a unified host computer provides a reference time to each sensor, so that each frame of sensor data is synchronized to a unified timestamp. Please refer to Fig.12 , because the timestamps of each sensor are different, the perception of each sensor is inaccurate: for example, if the timestamps of the camera and lidar are not synchronized, there may be a situation where the object in the image does not match the position of the object measured by the lidar, resulting in perception errors. Data fusion problems: Unmanned vehicles often use sensor data fusion to combine information from different sensors to obtain a more comprehensive understanding of the environment. Unsynchronized timestamps may cause data fusion problems because the data is not collected at the same time and the fusion algorithm must process data that is inconsistent in time. Reduced real-time performance: Unsynchronized timestamps of sensor data may cause delays in data processing and decision-making. Unmanned vehicles need to respond to environmental changes in real time, so timestamp synchronization is critical to real-time performance. Misleading decisions: The decision-making system of unmanned vehicles relies on accurate perception data, including vehicle position, obstacle detection, etc.

[0155] Data with different timestamps may mislead the decision-making system and cause the vehicle to make wrong decisions based on inaccurate information. Therefore, multi-sensor time synchronization is necessary. After the sensor timestamps are synchronized, the data between sensors can work together and merge more easily, improving the accuracy and reliability of perception and planning decisions.

[0156] Step 3.2: During the driving process of the unmanned vehicle, the camera collects images and the laser radar collects point cloud data, and then the images and point cloud data are preprocessed; image preprocessing includes measures such as adjusting image size, color balance, denoising, and edge detection. The goal of preprocessing is to optimize image quality and provide more accurate input for subsequent target detection. Point cloud data preprocessing includes removing outliers, filtering and denoising, and point cloud registration to reduce data uncertainty and noise.

[0157] Step 3.3, performing image registration on the preprocessed image and performing image fusion; performing point cloud registration on the preprocessed point cloud data and performing fusion on multiple point cloud data;

[0158] Step 3.4, perform target detection on the fused image and point cloud data;

[0159] For images, a target detection algorithm based on deep learning can be used. In this embodiment, YOLOv5s target detection is used to extract the category and location information of the target from the image. The core idea of ​​Yolo (You only look once) is grid division and target prediction: YOLO divides the input image into a fixed number of grid cells. Each grid cell is responsible for predicting whether there is a target in the image and the location information of the target. This step includes regressing the position coordinates of the center point of the target relative to the boundary of the grid cell, predicting the category information of the target and a confidence value, which represents the probability of the existence of the target and the accuracy of the target box.

[0160] Confidence and non-maximum suppression: For each target box prediction, YOLO calculates a confidence value to indicate whether the target box contains the object and the accuracy of the prediction. By setting the confidence threshold, target boxes with high confidence can be screened out to reduce false detections. Finally, non-maximum suppression (NMS) processing is used to remove overlapping bounding boxes to ensure that each target is detected only once, thereby obtaining the final target detection result.

[0161] The characteristics of the Yolov5s model in this embodiment are: 1. YOLOv5s introduces Mosaic data enhancement technology, which mixes multiple randomly cropped images into one, increases the diversity of training data, and helps improve the generalization ability of the model. This means that the model can better cope with targets of various scenes and object sizes. 2. YOLOv5 uses an adaptive anchor box calculation algorithm, which automatically calculates the appropriate anchor box size based on the target size in the training data set. This helps the model better adapt to targets of different sizes and improves the accuracy of detection. 3. The FPN and PAN structure are used on the Neck end: YOLOv5 uses FPN (Feature Pyramid Networks) and PAN (Pixel Aggregation Network) structures in feature extraction to further improve the feature extraction capabilities of the model. This means that the model can better capture the target information in the image, allowing it to perform better in target detection tasks. The specific flow chart of Yolo target detection is as follows Figure 6 shown.

[0162] Depend on Figure 6It can be seen that its target detection is divided into the following three steps: 1) YOLO divides the input image into SxS grids, and each grid is responsible for detecting objects whose center points fall within the grid. For each grid, B anchor boxes are indeed generated to predict the bounding box of the object. These anchor boxes are determined in pre-training and usually include multiple anchor boxes of different shapes and sizes. These anchor boxes are annotated relative to the upper left corner of each grid and used for position prediction. 2) Position and category prediction: In each grid, position prediction and category prediction are performed. Position prediction involves generating the bounding box positions of B anchor boxes and outputting the corresponding coordinates. Category prediction extracts features through convolutional neural networks, predicts C classification scores and corresponding confidence scores, which represent the probability that the object belongs to each category. 3) Data association and non-maximum suppression (NMS): For each anchor box, data association is performed in combination with the category confidence and category score, that is, the object category is associated with each anchor box.

[0163] Next, non-maximum suppression is used to eliminate multiple highly overlapping bounding boxes and retain the most likely detection results. NMS is a key step for removing redundant bounding boxes to ensure that each object is detected only once. 4) After data association and NMS processing, the final detection results are obtained. These results include the location bounding box of the object, the object category, and the confidence. It can be seen that the YOLO of the present invention can perform classification tasks and positioning tasks simultaneously, which can improve the speed of detection.

[0164] For point cloud data, the Euclidean clustering algorithm is used to divide the point cloud data into different target clusters. The core idea of ​​Euclidean clustering is to cluster points based on their spatial distance and feature attributes. Specifically, the Euclidean clustering algorithm calculates the Euclidean distance between each pair of points. If the distance between two points is less than a predefined threshold, the two points are classified into the same cluster. This process will be gradually expanded to classify all points within the threshold distance from existing clusters into the same cluster. In this way, different obstacles or targets in the point cloud data can be divided into different clusters. The specific process is as follows: Figure 7 The steps are as follows:

[0165] 1. Input point cloud segmentation: First, segment the point cloud data input by the lidar and divide it into small local point cloud sets. These point cloud sets may contain a single target or obstacle. 2. Point cloud clustering: For each point P in each point cloud set, calculate the Euclidean distance between point P and other points. If the Euclidean distance of a point is less than the set threshold, it is divided into the same cluster Q. 3. Gradual expansion of clusters: For each point in cluster Q (such as point P), use it as a new initial point and repeat the process of step 2. Add other points that meet the Euclidean distance threshold to cluster Q to gradually expand the size of the cluster. 4. Output clustering results: When the points in cluster Q no longer increase, output the clustering results. Each cluster contains points with similar attributes, which are clustered together to represent the same entity (such as a target or obstacle). LiDAR Euclidean clustering can effectively segment point cloud data into different obstacles or targets, providing more accurate data for subsequent target detection, tracking, and perception tasks.

[0166] Therefore, LiDAR Euclidean clustering can effectively divide point cloud data into different clusters through steps such as point cloud data segmentation, point cloud clustering division and gradual expansion of clusters, so as to better describe the targets and obstacles in the environment.

[0167] Step 3.5: Data-associate the object detection results of the image with the object detection results of the point cloud data to ensure that they correspond in time and space; this step requires the use of the sensor's timestamp and spatial coordinates to determine the matching of the same object in different sensor data.

[0168] Step 3.6: Target-level fusion is performed on the associated target data; target-level fusion is to merge the target information from different sensors, including the target category, position, speed, etc. The fusion method can be based on weighted average, probability fusion and other strategies to obtain more comprehensive and accurate target information.

[0169] Step 3.7: Output the fusion result, which includes the category, position, and speed of the target. The fusion result can be used for higher-level decision-making, control, and perception tasks, such as obstacle avoidance and path planning.

[0170] Therefore, the environmental perception of the present invention covers the sensor data acquisition, preprocessing and target-level fusion process. Through reasonable data association and fusion strategies, the fusion results can provide more accurate and comprehensive environmental perception, providing important support for the decision-making and control of unmanned vehicles.

[0171] Step 4: Unmanned vehicle A uses the processed environmental data, combined with its own linear speed and angular speed data, to perform local path planning, and sends the vehicle driving status data to the central cloud computing module; at the same time, when unmanned vehicle A drives to the area of ​​other edge computing cloud modules, the edge computing cloud modules interact with each other;

[0172] In this step, the dynamic window algorithm (DWA, Dynamic Window Approach) is used to plan the local path. Its basic idea is to implement it in three parts: model building, velocity sampling, and trajectory evaluation based on the robot's kinematic model and the motion characteristics of the kinematic model. Through the implementation of these three parts, the drive of the mobile robot can directly act on the velocity space (v, ω) and be restricted to the feasible dynamic space. At the same time, the unmanned vehicle samples in the velocity space, generates simulated trajectories, evaluates them according to the evaluation items, and finally selects the group with the highest score to obtain the local optimal path.

[0173] like Figure 8 As shown, the following steps are included:

[0174] Step 4.1: Establish the motion model of the unmanned vehicle: The unmanned vehicle moves within the time interval Δt, and the kinematic model is

[0175] x t+Δt =x t +v x Δtcosθ t -v y Δtsinθ t ;

[0176] y t+Δt =y t +v x Δtsinθ t +v y Δtcosθ t ;

[0177] θ t+Δt =θ t +ω t Δt;

[0178] In the formula, x t+Δt is the horizontal coordinate of the unmanned vehicle at time t+Δt, y t+Δt is the ordinate of the unmanned vehicle at time t+Δt, x t is the horizontal coordinate of the unmanned vehicle at time t, y t is the ordinate of the unmanned vehicle at time t, v x and v y is the linear velocity of the unmanned vehicle along the x and y axes at time t in the chassis coordinate system, θ t is the angle between the driverless car and the x-axis at time t, ω t is the angular velocity of the autonomous vehicle at time t;

[0179] Step 4.2, unmanned vehicle speed sampling: There are infinite sets of speed pairs (v, ω) in the two-dimensional speed space. According to the limitations of the unmanned vehicle itself and environmental factors, constraints are imposed to control the speed search space (v, ω) within a dynamic range, namely, the boundary limitation conditions:

[0180] v d ={(v,ω)|v∈[v min ,v max ]∧ω∈[ω min ,ω max ]};

[0181] Where: v m is the maximum and minimum speed limits of the unmanned vehicle, v max and v min are the maximum and minimum linear speeds that the unmanned vehicle can achieve, ω max and ω min They are the maximum angular velocity and minimum angular velocity that the unmanned vehicle can reach respectively;

[0182] Due to the maximum acceleration and deceleration limit imposed by the driving force of the unmanned vehicle, the speed space is further constrained by v d Indicates; Assume that the current linear velocity of the robot is v c , the current angular velocity is ω c , the time interval is Δt, then the achievable speed v at the next moment d for:

[0183]

[0184] In the formula, is the maximum acceleration; is the maximum deceleration; is the maximum angular acceleration; is the maximum angular deceleration;

[0185] In a time interval, the speed that the unmanned vehicle can reach is within a window, that is, the unmanned vehicle has a maximum acceleration limit:

[0186]

[0187] Where: dist(v,ω) represents the path length between the unmanned vehicle and the nearest obstacle;

[0188] Within a forward cycle of the unmanned vehicle, Δt, the effective speed set interval v is:

[0189] v=v m ∩v d ∩v a ;

[0190] Step 4.3: Use the evaluation function to evaluate the trajectories corresponding to the sampled groups of speeds to select the optimal trajectory; the criteria for trajectory evaluation are: accurately avoiding obstacles and approaching the target point in the shortest time; the evaluation function is:

[0191] G(v,ω)=α·head(v,ω)+β·dist(v,ω)+γ·vel(v,ω);

[0192] Where α, β and γ are the weight factors of each item, head(v, ω) is the azimuth evaluation function, which indicates the azimuth deviation between the direction of the unmanned vehicle and the end point of the entire path when it moves to the end point of a planned path at the current sampling speed. The smaller the value, the higher the evaluation score; dist(v, ω) is the safety factor evaluation function, which indicates the path length between the unmanned vehicle and the nearest obstacle; vel(v, ω) is the current speed evaluation function.

[0193] The aforementioned multi-vehicle cooperative control method based on edge and central cloud joint computing normalizes the azimuth evaluation function, the safety factor evaluation function, and the current speed evaluation function before speed evaluation. The normalization process is as follows:

[0194]

[0195]

[0196]

[0197] Where n is the trajectory in the dynamic window, and i is the trajectory that needs to be evaluated.

[0198] When the unmanned vehicle is driving along the globally planned path, DWA local path planning will be performed. The process is as follows: Step 1: The unmanned vehicle will receive the environmental perception data processed by its own sensors, the processed road conditions transmitted by the external module (edge ​​computing cloud / central cloud module), and the information perceived by other vehicles, and collect and send the vehicle's own linear velocity and angular velocity data to the external module. Step 2: The unmanned vehicle will generate a set of trajectory candidates based on the collected and transmitted data. Each trajectory is a possible movement path for the vehicle, including different combinations of speed and steering angle. Step 3: For each generated trajectory, the unmanned vehicle will use an evaluation function to evaluate its quality. The evaluation function takes into account multiple factors, such as the distance from obstacles, speed and direction deviation. These factors help determine which trajectories are safe and feasible. Step 4: Based on the results of the evaluation function, the unmanned vehicle will select the best trajectory. Then, it will drive along the selected best trajectory. At the same time, during the driving process of the unmanned vehicle, it will determine whether it has reached the end of the optimal trajectory. If it has reached the end of the optimal trajectory, it will continue to determine whether it has reached the end of the path. If it has reached the end, it will jump to step 5. If not, it will return to step 2. On the contrary, if it has not reached the end of the optimal trajectory, it will continue to move along the optimal trajectory. Step 5: The unmanned vehicle reaches the destination and ends this planning.

[0199] Step 5. Unmanned vehicle A adopts predictive control based on the MPC model to move along the reference point of the optimal trajectory; specifically, the three core elements of the MPC model predictive control include: 5.1. Prediction model: At the beginning of each control cycle, based on the current vehicle state and system model, predict the state and control input within a series of discrete time steps in the future. 5.2. Rolling optimization: For the above optimization problem, use numerical optimization methods (such as nonlinear programming or quadratic programming) to solve the optimal control input sequence. This process usually iterates multiple times in each control cycle to ensure that a smooth and feasible control input sequence is generated. 5.3. Feedback correction: At the beginning of each new control cycle, the actual state of the vehicle is measured, and the initial state is corrected based on the feedback information. This ensures that the MPC algorithm is robust to system disturbances and uncertainties. In this embodiment, if Fig. 9 As shown in the figure, the MPC model predictive control consists of an MPC controller, a state estimator, and a controlled object. The MPC controller uses the objective function, the prediction model, and the constraints to perform the optimal solution and obtain the optimal control sequence. The controlled object is controlled according to the input control quantity, and the current state quantity observation value is input into the state estimator. The obtained state quantity is input into the MPC controller for the next optimal solution, and this cycle repeats. It can be seen that the MPC algorithm can realize the trajectory tracking and control of the vehicle. This process includes key steps such as prediction model, rolling optimization, and feedback correction to achieve precise control in each control cycle.

[0200] Step 6: During the movement, determine whether there are no dynamic obstacles on the moving path. If so, return to step 3, reacquire the correlation information, update the global path and local path, and then re-perform motion control; if not, continue to determine whether it has moved to the end of the trajectory. If so, then determine whether it has reached the end. If it has not reached the end, return to step 3. If it has reached the end, output the path; otherwise, continue to predict control based on the MPC model and repeat step 5.

[0201] Based on the above method, the practical application of the present invention in roads is as follows Fig.10 As shown. Therefore, the data processing and decision-making mode based on the joint computing of edge and central cloud in the present invention not only improves the intelligence level and driving efficiency of the unmanned vehicle A, but also effectively reduces the time waste and ensures that the vehicle can travel more safely and efficiently in a complex road environment.

[0202] Embodiment 2: This embodiment provides a system for implementing a multi-vehicle collaborative control method based on edge and central cloud joint computing, such as Fig.11 As shown, it includes an unmanned driving module and an external module; the unmanned driving module is provided with an environment perception module, a decision-making and planning module, an on-board communication module and a control execution module; the external module is provided with an edge computing cloud module and a central cloud computing module; the decision-making and planning module is respectively connected to the environment perception module, the on-board communication module and the control execution module; the edge computing cloud module is connected to the central cloud computing module; the edge computing cloud module and the central cloud computing module are respectively connected to the decision-making and planning module and the environment perception module via the on-board communication module; the environment perception module is also connected to the control execution module.

[0203] Specifically, the environmental perception module is composed of multiple sensors, including cameras, lidar, GPS, and IMU sensors. These sensors work together to provide rich environmental information. Multiple cameras use the YOLOv5s target detection algorithm to identify pedestrians and vehicles in real time, while multiple lidars use the Euclidean clustering algorithm to convert point cloud data into clustering boxes of pedestrians and vehicles. The results of target detection by these two types of sensors are fused and sent to the decision-making and planning module for planning. The target detection results need to be synchronized in time and space to ensure data consistency.

[0204] Decision-making and planning module: This module receives data sent from the perception and recognition module and data such as the environment and traffic conditions transmitted by external modules through the 5G communication of the vehicle communication network. Combined with this information, the decision-making and planning module makes advanced decisions, such as overtaking, obstacle avoidance, parking, etc., and performs path planning to select the optimal driving route. In addition, considering the dynamic environmental changes, the module may use prediction algorithms to predict the behavior of surrounding vehicles and pedestrians to make more intelligent decisions.

[0205] In-vehicle communication module: As a bridge between the unmanned vehicle and the outside world, the in-vehicle communication module realizes high-speed data interaction between the vehicle and the edge computing cloud module and the vehicle and the central cloud computing module. As a result, the unmanned vehicle can obtain the latest map data, traffic information, environmental perception results, and real-time decision results. At the same time, it transmits its own status and updated map data to the edge computing cloud module or the central cloud computing module to achieve global information sharing and collaboration.

[0206] Control execution module: This module converts the high-level decisions and path planning results of the decision-making and planning module into specific vehicle actions. Through the vehicle's control system, it controls the vehicle's speed, steering, acceleration, etc. This module is not only responsible for path tracking, but also needs to perceive the vehicle's surrounding environment in real time, adjust the vehicle's actions, and ensure driving safety and stability.

[0207] Edge computing cloud module: The edge computing cloud module is located outside the unmanned vehicle. It is a distributed computing model that places computing and data storage at the edge of the network, which can improve the speed and efficiency of data processing and reduce the cost and delay of data transmission. In an unmanned vehicle, the unmanned vehicle sends the real-time perception data of its sensors and other perception data sent by the vehicle communication module and the Internet of Vehicles cloud platform in the cloud computing module to the edge computing cloud module for perception calculation. The edge computing cloud module will send the results of its perception calculation to the unmanned vehicle to improve the vehicle's environmental perception ability. At the same time, the edge computing cloud module will combine the results of perception calculation, the results of positioning calculation, and information data such as high-definition maps to make real-time decision planning and send it to the unmanned vehicle, thereby optimizing the decision of unmanned driving. More importantly, the real-time decision made by the unmanned vehicle is based on various evaluation indicators, such as safety, time and other factors. The real-time decision with the highest evaluation is selected from the real-time decision of the unmanned vehicle itself and the real-time decision sent by the edge computing cloud module. In addition, the edge computing cloud module will also send the latest map data, traffic conditions and other data to the central cloud computing module to update various data.

[0208] Central cloud computing module: The central cloud computing module contains various cloud platforms, such as the Internet of Vehicles cloud platform, transportation management cloud platform, map cloud platform, etc. This module integrates, processes, analyzes and supports decision-making for various types of data. At the same time, it receives the status information and location information of the unmanned vehicle and sends the perception information and location information of other vehicles to the unmanned vehicle to ensure that the vehicle can travel safely and efficiently. This module is usually supported by large-scale computing and storage infrastructure to cope with complex tasks and real-time needs.

[0209] In summary, the various modules of the unmanned vehicle system work closely together to achieve autonomous driving of the vehicle. The perception and recognition module provides environmental information, the decision-making and planning module formulates strategies, the vehicle communication module realizes data interaction, the control execution module converts decisions into actions, and the central cloud computing module and edge computing module of the external module provide the data required by the unmanned vehicle. The collaborative mode of interaction between the internal and external modules of the unmanned vehicle brings the unmanned vehicle into a more intelligent and safe driving era.

[0210] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0211] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0212] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0213] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A multi-vehicle cooperative control method based on edge and central cloud joint computing, characterized in that: The multi-vehicle cooperative control method based on edge and central cloud joint computing includes the following steps: Step 1: When any unmanned vehicle A starts, its target destination is given. The central cloud computing module is used to control the unmanned vehicle A and receive vehicle data. At the same time, the central cloud computing module sends the map and traffic status data to the edge computing cloud module a corresponding to the unmanned vehicle A. The edge computing cloud module a then sends the processed map and traffic status data to the unmanned vehicle A. Step 2: Unmanned vehicle A uses a global path planning algorithm to combine the map processed by edge computing cloud module a, traffic data, and the location data of the unmanned vehicle itself to generate a global path that will guide unmanned vehicle A from the starting point to the target location. Step 3: Based on the request from unmanned vehicle A, the central cloud computing module calculates the correlation between unmanned vehicle A and other unmanned vehicles, and sends the environmental data sensed by the highly correlated unmanned vehicles to the edge computing cloud module a. At the same time, unmanned vehicle A sends its own perceived environmental data to the edge computing cloud module a. After receiving the environmental data, the edge computing cloud module a processes and updates it and sends it to the central cloud computing module. Step 4: Unmanned vehicle A uses the processed environmental data, combined with its own linear velocity and angular velocity data, to perform local path planning and send the vehicle's driving status data to the central cloud computing module. At the same time, when unmanned vehicle A drives into the area of ​​another edge computing cloud module, the edge computing cloud modules exchange data with each other. Step 5: Unmanned vehicle A uses MPC model-based predictive control to move along the reference point of the optimal trajectory; Step 6: During the movement process, determine whether there are no dynamic obstacles on the moving path. If so, return to step 3, reacquire the correlation information, update the global path and local path, and then re-perform motion control; if not, continue to determine whether it has moved to the end point of the trajectory. If so, then determine whether it has reached the end point. If it has not reached the end point, return to step 3. If it has reached the end point, output the path; otherwise, continue to predict control based on the MPC model and repeat step 5.

2. The multi-vehicle cooperative control method based on edge and central cloud joint computing according to claim 1 is characterized in that: The global path planning algorithm is an improved A-star algorithm, and the steps of performing global path planning by the improved A-star algorithm are as follows: Step 2.1: Autonomous vehicle A receives map and traffic data, and obtains its own location data. Step 2.2: Rasterize and dilate the map, then initialize the open list and closed list in the A-star algorithm, and put the starting point into the open list; Step 2.3: Traverse the open list to find the node corresponding to the minimum evaluation function, put it into the closed list, and set it as the current node N; Step 2.4: Determine whether the current node N is the end point. If so, output the planned path and jump to step 7. Otherwise, if not, proceed to step 5. Step 2.5: Expand the current node N according to the corresponding expansion method, and put the expanded nodes that are not in the open list into the open list; Step 2.6: Check whether the open list is empty. If it is empty, the path planning fails. Otherwise, if it is not empty, return to step 3. Step 2.7: Use a second-order Bezier curve to smooth the planned path to obtain the optimal path.

3. The multi-vehicle cooperative control method based on edge and central cloud joint computing according to claim 1 is characterized in that: The correlation consists of position distance correlation and speed correlation; The position distance correlation dis(A, B) is expressed as follows: Where: x_A and y_A represent the position coordinates of vehicle A, x_B and y_B represent the position coordinates of vehicle B; The velocity correlation vel(A,B) is expressed as follows: vel(A,B)=abs(v_A-v_B) Where: v_A represents the speed of vehicle A, v_B represents the speed of vehicle B; abs() represents the absolute value function; According to the position distance and speed correlation formula, the correlation formula cor(A,B) is defined as follows: cor(A,B)=k1*dis(A,B)+k2*vel(A,B) Among them, k1 and k2 are weight factors used to balance the influence of different factors.

4. The multi-vehicle cooperative control method based on edge and central cloud joint computing according to claim 1 is characterized in that: The steps to obtain the perceived environmental data are as follows: Step 3.1: Synchronize the multiple cameras and lidars of the autonomous vehicle in time and space to ensure that data is captured at the same timestamp. Step 3.2: While the unmanned vehicle is moving, the camera collects images and the lidar collects point cloud data, and then the images and point cloud data are preprocessed; Step 3.3, perform image registration on the preprocessed images and fuse the images; Perform point cloud registration on the pre-processed point cloud data and fuse multiple point cloud data; Step 3.4: Perform target detection on the fused image and point cloud data; Step 3.5: Associate the object detection results of the image with those of the point cloud data to ensure that they correspond in time and space. Step 3.6: Perform target-level fusion on the associated target data; Step 3.7: Output the fusion result, which includes the category, position, and speed of the target.

5. The multi-vehicle cooperative control method based on edge and central cloud joint computing according to claim 1 is characterized in that: The steps of performing local path planning are specifically as follows: Step 4.1: Establish the motion model of the unmanned vehicle: The unmanned vehicle moves within the time interval Δt, and the kinematic model is: x t+Δt =x t +v x Δt cosθ t -v y Δt sinθ t ; y t+Δt =y t +v x Δt sinθ t +v y Δt cosθ t ; i t+Δt =θ t +oh t Δt; Where x t+Δt is the horizontal coordinate of the unmanned vehicle at time t+Δt, y t+Δt is the vertical coordinate of the unmanned vehicle at time t+Δt, x t is the horizontal coordinate of the autonomous vehicle at time t, y t is the vertical coordinate of the autonomous vehicle at time t, v x and v y is the linear velocity of the unmanned vehicle along the x and y axes at time t in the chassis coordinate system, θ t is the angle between the autonomous vehicle and the x-axis at time t, ω t is the angular velocity of the autonomous vehicle at time t; Step 4.2, Unmanned Vehicle Speed ​​Sampling: There are infinite sets of speed pairs (v, ω) in the two-dimensional speed space. Based on the limitations of the unmanned vehicle itself and environmental factors, constraints are imposed to control the speed search space (v, ω) within a dynamic range, i.e., the boundary constraint conditions: v d ={(v,ω)|v∈[v min ,v max ]^ω∈[ω min ,ω max ]}; Where: v m is the maximum speed and minimum speed limit range of the unmanned vehicle, v max and v min are the maximum and minimum linear speeds that the unmanned vehicle can achieve, ω max and ω min are the maximum and minimum angular velocities that the unmanned vehicle can reach respectively; Due to the maximum acceleration and deceleration limit imposed by the driving force of the unmanned vehicle, the speed space is further constrained by using v d Indicates; let the current linear velocity of the robot be v c , the current angular velocity is ω c , the time interval is Δt, then the achievable speed v at the next moment is d for: Where, is the maximum acceleration; is the maximum deceleration; is the maximum angular acceleration; is the maximum angular deceleration; In a time interval, the speed that the autonomous vehicle can reach is within a window, that is, the autonomous vehicle has a maximum acceleration limit: Where: dist(v,ω) represents the path length between the autonomous vehicle and the nearest obstacle; During the unmanned vehicle's forward cycle of one Δt, the effective speed set interval v is: v=v m ∩v d ∩v a ; Step 4.3: Use an evaluation function to evaluate the trajectories corresponding to the sampled speed groups to select the optimal trajectory. The trajectory evaluation criteria are: accurately avoiding obstacles and approaching the target point in the shortest time. The evaluation function is: G(v,ω)=α·head(v,ω)+β·dist(v,ω)+γ·vel(v,ω); Where α, β, and γ are the weight factors of each item; head(v, ω) is the azimuth evaluation function, which indicates the azimuth deviation between the direction of the unmanned vehicle and the end point of the entire path when moving to the end point of a planned path at the current sampling speed. The smaller the value, the higher the evaluation score; dist(v, ω) is the safety factor evaluation function, which indicates the path length between the unmanned vehicle and the nearest obstacle; vel(v, ω) is the current speed evaluation function.

6. The multi-vehicle cooperative control method based on edge and central cloud joint computing according to claim 5 is characterized in that: Before speed evaluation, the azimuth angle evaluation function, safety factor evaluation function and current speed evaluation function are normalized. The normalization process is as follows: Where n is the trajectory in the dynamic window, and i is the trajectory that needs to be evaluated.

7. The multi-vehicle cooperative control method based on edge and central cloud joint computing according to claim 1 is characterized in that: The steps of the MPC model predictive control are specifically as follows: Step 5.

1. Prediction model: At the beginning of each control cycle, based on the current vehicle state and the system model, predict the state and control inputs in a series of discrete time steps in the future; Step 5.2, rolling optimization: use numerical optimization methods to solve and obtain the optimal control input sequence; Step 5.3: Feedback correction: At the beginning of each new control cycle, the actual state of the vehicle is measured and the initial state is corrected based on the feedback information.

8. A system for multi-vehicle cooperative control method based on edge and central cloud joint computing, characterized in that: The system of the multi-vehicle collaborative control method based on edge and central cloud joint computing includes an unmanned driving module and an external module; the unmanned driving module is provided with an environmental perception module, a decision-making and planning module, an on-board communication module and a control execution module; the external module is provided with an edge computing cloud module and a central cloud computing module; the decision-making and planning module is respectively connected to the environmental perception module, the on-board communication module and the control execution module; the edge computing cloud module is connected to the central cloud computing module; the edge computing cloud module and the central cloud computing module are respectively connected to the decision-making and planning module and the environmental perception module via the on-board communication module; the environmental perception module is also connected to the control execution module.

9. The system of multi-vehicle cooperative control method based on edge and central cloud joint computing according to claim 8 is characterized in that: The environmental perception module includes multiple sensors, including cameras, laser radars, GPS and IMU sensors; The decision-making and planning module is used to receive data sent from the environmental perception and recognition module and environmental and traffic condition data transmitted from the external module through the vehicle communication module, and perform global path planning and local path planning; The vehicle communication module is used to realize high-speed data interaction between the unmanned vehicle and the edge computing cloud module and the central cloud computing module; The control execution module is used to convert the decision and path planning results of the decision planning module into specific vehicle actions; The edge computing cloud module is located outside the unmanned vehicle and is a distributed computing model that places computing and data storage at the edge of the network; The central cloud computing module is equipped with multiple cloud platforms, including the Internet of Vehicles cloud platform, the transportation management cloud platform and the map cloud platform.

Images