Unmanned driving control system

Abstract

The invention, which relates to the technical field of unmanned driving, discloses an unmanned driving control system comprising the following components: an environment sensing layer, a positioning mapping layer, a path planning layer, a decision control layer and a bottom execution layer. According to the invention, the improved YOLOv7-tiny target detection algorithm and the DeepSort target tracking algorithm are adopted, and laser radar point cloud processing is combined, so that accurate identification and tracking of track markers such as cone buckets are realized, the real-time performance and accuracy of target detection are improved, the robustness of environmental perception is further enhanced through a data fusion technology, and the method has the advantages of being high in practicability and easy to popularize. Meanwhile, on the path planning and decision control level, a graph search tree and a B spline curve are used for optimizing the path, and a Stanley / MPC hybrid transverse control algorithm is combined, so that the racing car can still keep stable driving track and speed in high-speed driving and complex racing track environments, and the stability and accuracy of an unmanned driving system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned driving technology, in particular to an unmanned driving control system. Background Art

[0002] Today, as driverless technology develops rapidly, driverless control systems, as its core component, have a performance that directly affects the safety, efficiency and intelligence level of driverless vehicles. Especially in high-level competitions such as the China University Driverless Formula Competition, the requirements for driverless systems are more stringent. Not only do they need to have high-precision environmental perception capabilities and fast and accurate path planning capabilities, but they also need to be able to cope with complex and changing competition environments and achieve stable and reliable vehicle control.

[0003] In the design of traditional driverless racing cars, the environmental perception layer mainly obtains surrounding environment information through sensors such as lidar and cameras, but there are problems such as low data fusion efficiency, poor real-time target detection and tracking, and susceptibility to environmental interference. Although lidar has high precision, it has a large data processing volume and is easily affected by bad weather such as rain and fog. Although the camera has a wide viewing angle, the detection effect is easily affected when the lighting changes greatly or there are obstructions. In addition, traditional path planning methods are mostly based on static maps or simple dynamic obstacle avoidance strategies, which are difficult to cope with complex and changeable track environments, and path optimization algorithms are insufficient in real-time, accuracy and robustness. These problems limit the application effect of driverless technology in complex scenarios.

[0004] In view of the above problems, it is necessary to optimize the existing unmanned driving control system. By introducing the improved YOLOv7-tiny target detection algorithm and DeepSort target tracking module, combined with the lidar point cloud processing and fusion algorithm, the real-time and accuracy of environmental perception can be improved. At the same time, the laser SLAM algorithm is used for real-time map construction and positioning, and the Delaunay triangulation and graph search tree algorithm are combined for path planning to ensure the smoothness and rationality of the path. Therefore, it is of great significance to develop an unmanned driving control system that can comprehensively realize the above characteristics. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art and provide an unmanned driving control system, which can significantly improve the real-time and accuracy of environmental perception by introducing an improved YOLOv7-tiny target detection algorithm and a DeepSort target tracking module, combined with a lidar point cloud processing and fusion algorithm, while using a laser SLAM algorithm for real-time map construction and positioning, combining Delaunay triangulation with a graph search tree algorithm for path planning, ensuring the smoothness and rationality of the path, and at the decision-making control layer, through a hybrid lateral control algorithm and a piecewise speed planning method, achieving precise control of vehicle motion, in addition, by building a perception experimental platform and adopting an advanced underlying control system, the overall performance and reliability of the system are further improved.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: an unmanned driving control system, the system includes the following components: an environment perception layer, a positioning construction layer, a path planning layer, a decision control layer and a bottom execution layer;

[0007] The environmental perception layer includes an image processing module, a point cloud processing module and a data fusion module, and is equipped with a high-precision laser radar and a camera. Through the improved YOLOv7-tiny target detection algorithm and DeepSort target tracking algorithm, it can realize real-time detection and tracking of targets such as cones and barrels;

[0008] The positioning and construction layer adopts the SC-LeGO-LOAM laser SLAM algorithm to build a real-time map through the scanning data of the laser radar and perform precise positioning;

[0009] The path planning layer discretizes the global map using Delaunay triangulation, generates the target path through a graph search tree algorithm, and uses a cubic B-spline curve for path fitting;

[0010] The decision control layer uses a hybrid lateral control algorithm based on DDPG and a three-point curvature velocity planning method to control the front wheel steering angle and acceleration and deceleration of the vehicle according to the path information and vehicle status;

[0011] The bottom execution layer receives control signals through the CAN network and precisely controls the steering system, the braking system and the driving system.

[0012] Furthermore, the environmental perception layer includes an image processing module, a point cloud processing module and a data fusion module, wherein the image processing module obtains a 2D detection frame of the target by realizing target detection and tracking, the point cloud processing module realizes target clustering, extracts target point cloud and three-dimensional coordinate information, and the data fusion module parses the coordinates, color and confidence information of the cone barrel in the current field of view by fusing vision and lidar data, and provides it to the path planning layer for use.

[0013] Furthermore, the image processing module in the environment perception layer uses an improved YOLOv7-tiny target detection algorithm to achieve target detection and tracking. Specifically, image data containing a target object is received and the image data is processed using a model. The model includes a feature extraction layer, a target classification layer, and a bounding box regression layer. In the target classification layer, the classification loss L is calculated by a cross entropy loss function. cls , in the bounding box regression layer, the bounding box regression loss L is calculated by the squared error loss function bbox , in order to optimize the model's ability to predict the target bounding box, the classification loss and the bounding box regression loss are weighted summed to obtain the total loss, which is calculated as: L total =αL cls +βL bbox , where α and β are hyperparameters, and the model is trained and optimized based on the total loss. Based on the real box and the predicted box, the DeepSort target tracking algorithm is used to calculate the CIoU value between each detection box and each predicted box, and the allocation cost matrix is ​​constructed. The calculation formula of CIoU is: Among them, IoU is the intersection over union ratio, b and b gt are the center points of the predicted box and the real box respectively, ρ 2 (b,b gt ) is the square of the Euclidean distance between the center points, c is the diagonal distance of the minimum closure area, v is the aspect ratio consistency measure, α is the trade-off parameter, and the Hungarian algorithm is used to match the detection box with the prediction box based on the above allocation cost matrix, and a cascade matching mechanism is introduced to improve the matching reliability. Finally, the Kalman filter is updated with the successfully associated detection box, and the updated bounding box is passed to the next frame as the tracking box. By combining the recognition result with the target tracking result, the recognition of the cone bucket color can be visualized based on the tracking count. At the same time, when the target detector fails to detect the target, the position of the cone bucket target can still be predicted.

[0014] Furthermore, the point cloud processing module in the environmental perception layer filters out the point clouds outside the rectangular area where the cones are distributed according to the situation of the stadium, and uses the RANSAC algorithm to select the point cloud with the largest number of random fittings in the iterative process according to the set threshold as the extracted ground point cloud, filter out the ground noise points, and use radius filtering to filter out the outliers to obtain the filtered point cloud, and then use Euclidean clustering to obtain the clustered point cloud cluster according to the set clustering sphere radius and clustering point cloud threshold, and limit the height and width of the clustered point cloud cluster according to the cone characteristics, filter out obstacles outside the cones, screen out the cones that meet the feature information, and send the extracted cone point cloud and its position information to the data fusion module.

[0015] Furthermore, the environmental perception layer data fusion module receives the cone bucket point cloud cluster and its three-dimensional bounding box obtained by the point cloud clustering module and the image bounding box obtained by the image processing module, projects the extreme value of the three-dimensional bounding box to the pixel plane, obtains the two-dimensional bounding box, and then uses the Hungarian algorithm to perform matching calculations, and uses the intersection and union ratio of the distances between the three-dimensional bounding box and the two-dimensional bounding box as the allocation cost matrix to complete the matching of the image and point cloud detection results, thereby realizing the fusion association of the cone bucket position information and the category information.

[0016] Furthermore, the positioning construction layer adopts the SC-LeGO-LOAM laser SLAM algorithm to construct a real-time map through the scanning data of the laser radar and perform precise positioning. Specifically, the laser radar scans the surrounding environment and generates point cloud data, and extracts feature points from the point cloud data. The real-time map is constructed through the feature matching algorithm, and the map and the current laser scanning data are used to perform precise positioning through the laser SLAM algorithm. Suppose the laser radar scanning data is S={p1,p2,...,p N}, where p i is a point in the scan with polar coordinates (r i ,θ i ), divide θ into M intervals of equal width, the width of each interval is For each interval j, calculate the maximum and minimum radius r of the points that fall within the interval max,j , r min,j , the Scan Context algorithm is used to process the scan data. The Scan Context histogram H is an M-dimensional vector, where H[j] = (r min,j ,r max,j ), loops are identified by comparing the similarity between the current scan and the historical scan. Suppose the two ScanContext histograms are H1 and H2, then the similarity D between them can be obtained by calculating the Euclidean distance between the corresponding intervals of the two histograms, that is,

[0017] Furthermore, the path planning layer discretizes the global map into a grid composed of points and lines using the Delaunay triangulation algorithm. In the discretized grid, the graph search tree algorithm is used to generate the optimal path from the start point to the end point based on the actual cost and estimated cost between nodes. The generated path points are used as control points and fitted using a cubic B-spline curve to obtain a smooth path with continuous curvature. The expression of the cubic B-spline curve is: Where P(t) is the position of any point on the path, B j,3 (t) is the cubic B-spline basis function, P i+j It is the control point.

[0018] Furthermore, the decision control layer receives the path information generated by the path planning layer and the current state information of the vehicle, and adopts a hybrid lateral control algorithm based on DDPG according to the current speed of the vehicle. The Stanley algorithm is used to calculate the front wheel angle in the low-speed range, and the MPC algorithm is used to calculate the front wheel angle in the high-speed range. The weights of the Stanley and MPC algorithms are dynamically adjusted by the DDPG algorithm in the transition range to achieve a smooth transition. By adopting a three-point curvature speed planning method, the speed of the vehicle is calculated and planned in segments according to the curvature change of the path, so that the vehicle can accelerate and decelerate smoothly during driving, and the calculated front wheel angle and speed are used as control instructions, which are sent to the underlying actuator through the vehicle control system to achieve control of the vehicle's front wheel angle and acceleration and deceleration.

[0019] Furthermore, the underlying execution layer adopts a PID control algorithm, and according to the error between the expected angle and the actual angle and the rate of change thereof, calculates the output angle and controls the steering actuator by adjusting the proportional, integral, and differential gains. A fuzzy PID control algorithm is adopted, with the actual speed, expected speed, speed error and its rate of change as input, and the PID parameters are dynamically adjusted through fuzzy reasoning to achieve precise control of the vehicle braking system.

[0020] Compared with the prior art, this unmanned driving control system has the following beneficial effects:

[0021] 1. The present invention adopts the improved YOLOv7-tiny target detection algorithm and DeepSort target tracking algorithm, combined with lidar point cloud processing, to achieve accurate recognition and tracking of track markers such as cones, which not only improves the real-time and accuracy of target detection, but also further enhances the robustness of environmental perception through data fusion technology. At the same time, at the path planning and decision control level, the graph search tree and B-spline curve are used to optimize the path, combined with the Stanley / MPC hybrid lateral control algorithm, so that the car can still maintain a stable driving trajectory and speed in high-speed driving and complex track environments, significantly improving the stability and accuracy of the unmanned driving system.

[0022] 2. The present invention introduces laser SLAM technology, and uses the SC-LeGO-LOAM algorithm to build a track environment map in real time and perform precise positioning, providing reliable navigation information for the car. It not only improves the car's autonomous navigation capability in unknown or complex environments, but also enhances the system's adaptability and robustness in dealing with track changes and emergencies. In addition, by introducing the DDPG reinforcement learning method, the weight switching of the Stanley and MPC control algorithms is optimized, achieving a smooth transition of control strategies in different speed ranges, further improving the intelligence level and overall performance of the unmanned driving system.

[0023] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1 A schematic diagram of an unmanned driving control system process;

[0026] Figure 2 It is a flow chart of an unmanned driving control system;

[0027] Figure 3 The figure is a point cloud processing flow chart for an unmanned driving control system. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] Embodiment 1

[0030] This embodiment describes in detail a specific application process of an unmanned driving control system, which includes an environment perception layer, a positioning and construction layer, a path planning layer, a decision control layer, and an underlying execution layer. The layers work together through efficient data transmission and processing mechanisms to achieve autonomous navigation and precise control of a racing car.

[0031] The environmental perception layer includes sensors and control hardware, image processing modules, point cloud processing modules and data fusion modules. The sensor uses the Beike Tianhui 32-line laser radar as the main sensor to collect point cloud data of targets such as cones and barrels. At the same time, it is equipped with a Daheng Industrial wide-angle camera to collect real-time images with a higher acquisition frequency and viewing angle. The Guangzhou Daoyuan GNSS / IMU combined navigation is selected to obtain high-precision vehicle position, heading angle and speed information. The industrial computer uses the VIOORBRAV-7601-S001 model, which has powerful computing power and uses an independent graphics card for GPU accelerated computing. The vehicle controller uses Huahai Technology Rapid ECU-U2 supports MATLAB / Simulink / Stateflow / ECUCoder, which is convenient for rapid prototyping and algorithm verification. The image processing module adopts the improved YOLOv7-tiny target detection algorithm and combines it with the DeepSort target tracking module to realize real-time detection and tracking of cone barrels. The point cloud processing module processes the original point cloud through methods such as through filtering, RANSAC algorithm, radius filtering and Euclidean clustering to extract the cone barrel point cloud information. The data fusion module fuses the image and point cloud data through joint calibration and time synchronization to obtain the precise location information and category information of the cone barrel.

[0032] The positioning and layer construction adopts the SC-LeGO-LOAM laser SLAM algorithm, builds a real-time map through the laser radar scanning data, and performs precise positioning. Specifically, the laser radar scans the surrounding environment and generates point cloud data, and extracts feature points from the point cloud data. The real-time map is built through the feature matching algorithm. The map and the current laser scanning data are used to perform precise positioning through the laser SLAM algorithm. Suppose the laser radar scanning data is S = {p1, p2, ..., p N}, where p i is a point in the scan with polar coordinates (r i ,θ i ), divide θ into M intervals of equal width, the width of each interval is For each interval j, calculate the maximum and minimum radius r of the points that fall within the interval max,j , r min,j , the Scan Context algorithm is used to process the scan data. The Scan Context histogram H is an M-dimensional vector, where H[j] = (r min,j ,r max,j ), loops are identified by comparing the similarity between the current scan and the historical scan. Suppose the two Scan Context histograms are H1 and H2, then the similarity D between them can be obtained by calculating the Euclidean distance between the corresponding intervals of the two histograms, that is, In practical applications, by continuously performing lidar scanning, real-time point cloud data is obtained, and the data is processed and analyzed. The extracted feature points and the constructed real-time map are combined with the current laser scanning data to perform precise positioning through the laser SLAM algorithm. At the same time, by identifying loops, the map can be continuously optimized and updated to improve the accuracy and completeness of the map. In this way, the unmanned driving system can perceive the surrounding environment more accurately, achieve precise positioning and navigation, and ensure safe and stable driving.

[0033] The path planning layer uses the Delaunay triangulation algorithm to discretize the global map into a grid composed of points and lines. The Delaunay triangulated grid is generated between cones by an incremental algorithm. The advantage of the incremental algorithm is that the identified cones can be dynamically added as grid nodes to update the grid as the car moves. The triangulation discretizes the global map so that there are only two elements in the global map: points and lines. This is a typical graph search problem, so the graph search tree algorithm is selected for path generation. First, nodes are sampled in the Delaunay triangulated grid space. The filter removes severely non-compliant nodes according to the rules. Then, the N-fork decision tree method is used to generate the optimal path from the start point to the end point based on the actual cost and estimated cost between nodes. The generated path points are used as control points and fitted using cubic B-spline curves to obtain a smooth path with continuous curvature. The expression of the cubic B-spline curve is: Where P(t) is the position of any point on the path, B j,3 (t) is the cubic B-spline basis function, P i+j It is the control point.

[0034] The decision control layer receives the path information generated by the path planning layer and the current state information of the vehicle. According to the current speed of the vehicle, the hybrid lateral control algorithm based on DDPG is adopted. The Stanley algorithm is used to calculate the front wheel angle in the low-speed range. The calculation formula is: Among them, K p and K d are proportional and differential gains respectively, θ is the deviation angle between the vehicle's current heading and the target path. The MPC algorithm is used to calculate the front wheel steering angle in the high-speed section, and the DDPG algorithm is used to dynamically adjust the weights of the Stanley and MPC algorithms in the transition section to achieve a smooth transition. The calculation formula is: The constraints are: stx k+1 =f(x k ,u k ),x0=x initial ,x k ∈X,u k∈U, where N is the prediction horizon, Q and R are weight matrices, f is the vehicle dynamics model, X and U are the constraint sets of state and control inputs, respectively, and x k and x ref,k They represent the state parameters of different path points respectively. By adopting the three-point curvature speed planning method, the vehicle speed is calculated and planned in segments according to the curvature change of the path, so that the vehicle can accelerate and decelerate smoothly during driving. The calculated front wheel steering angle and speed are used as control instructions and sent to the underlying actuator through the vehicle control system to realize the control of the vehicle's front wheel steering angle and acceleration and deceleration.

[0035] The bottom execution layer adopts PID control algorithm, which calculates the output angle and controls the steering actuator by adjusting the proportional, integral and differential gains according to the error between the expected angle and the actual angle and its rate of change. Where, e(t) is the error between the actual angle and the expected angle, K p , K i , K d They are proportional, integral and differential gains respectively. Fuzzy PID control algorithm is used at the same time. The actual speed, expected speed, speed error and its change rate are input. The calculation formula is: Among them, ΔK p ,ΔK i ,ΔK d They are the adjustment amounts of proportional, integral and differential gains respectively. The PID parameters are adjusted dynamically through fuzzy reasoning to achieve precise control of the vehicle braking system.

[0036] In summary, this embodiment achieves high-precision and high-stability unmanned driving in complex environments by integrating advanced sensor technology, optimization algorithms and precise control strategies.

[0037] Embodiment 2

[0038] Based on the first embodiment, this embodiment describes in detail a specific application of an environment perception layer in an unmanned driving control system.

[0039] The environmental perception layer includes an image processing module, a point cloud processing module and a data fusion module. The image processing module obtains the 2D detection frame of the target by realizing target detection and tracking. The point cloud processing module realizes target clustering and extracts the target point cloud and 3D coordinate information. The data fusion module parses the coordinates, color and confidence information of the cone barrel in the current field of view by fusing vision and lidar data, and provides it to the path planning layer.

[0040] The sensor uses Beike Tianhui 32-line laser radar as the main sensor to collect point cloud data of targets such as cones and barrels. At the same time, it is equipped with Daheng Industrial wide-angle camera to collect real-time images with a higher acquisition frequency and viewing angle. Guangzhou Daoyuan GNSS / IMU combined navigation is selected to obtain high-precision vehicle position, heading angle and speed information. The industrial computer uses VIOORBRAV-7601-S001 model, which has powerful computing power and uses an independent graphics card for GPU accelerated computing. The vehicle controller uses Huahai Technology RapidECU-U2, which supports MATLAB / Simulink / Stateflow / ECUCoder, which is convenient for rapid prototype development and algorithm verification.

[0041] The image processing module uses the improved YOLOv7-tiny target detection algorithm to achieve target detection and tracking. Specifically, it receives image data containing the target object and processes the image data using a model. The model includes a feature extraction layer, a target classification layer, and a bounding box regression layer. In the target classification layer, the classification loss is calculated using the cross entropy loss function, and its calculation formula is: Among them, p i is the probability of each type, y i is the image information, L cls is the classification loss, C is the category type, and in the bounding box regression layer, the bounding box regression loss L is calculated by the squared error loss function bbox , in order to optimize the model's ability to predict the target bounding box, the calculation formula is: Among them, L bbox is the bounding box regression loss, is the true bounding box, t i is the predicted bounding box, λ i L is the weight of different parameters, x and y are the center coordinate offsets, w and h are the width and height scaling, and the classification loss and bounding box regression loss are weighted summed to get the total loss, which is calculated as: total =αL cls +βL bbox , where α and β are hyperparameters, and the model is trained and optimized based on the total loss. Based on the real box and the predicted box, the DeepSort target tracking algorithm is used to calculate the CIoU value between each detection box and each predicted box, and the allocation cost matrix is ​​constructed. The calculation formula of CIoU is: Among them, IoU is the intersection over union ratio, b and b gt are the center points of the predicted box and the real box respectively, ρ 2 (b,b gt) is the square of the Euclidean distance between the center points, c is the diagonal distance of the minimum closure area, v is the aspect ratio consistency measure, α is the trade-off parameter, and the Hungarian algorithm is used to match the detection box with the prediction box based on the above allocation cost matrix, and a cascade matching mechanism is introduced to improve the matching reliability. Finally, the Kalman filter is updated with the successfully associated detection box, and the updated bounding box is passed to the next frame as the tracking box. By combining the recognition result with the target tracking result, the recognition of the cone bucket color can be visualized based on the tracking count. At the same time, when the target detector fails to detect the target, the position of the cone bucket target can still be predicted.

[0042] The point cloud processing module in the perception layer filters out the point clouds outside the rectangular area of ​​cone distribution according to the situation of the stadium, and then uses the RANSAC algorithm to select the point cloud with the largest number of random fittings in the iterative process as the extracted ground point cloud according to the set threshold. Specifically, three points are randomly selected from the point cloud data, and the plane equation is calculated using these three points. The plane equation can be expressed as: Ax+By+Cz+D=0, where A, B, and C are the normal vectors of the plane, and D is the distance from the plane to the origin. For each point in the point cloud, its distance to the plane is calculated. If the distance is less than a certain threshold t, the point is considered to be an inner point, and all points are used. The inliers are used to recalculate the plane equation to improve the accuracy of the model. The above steps are repeated for multiple times. The number of inliers is calculated in each iteration. The iteration with the largest number of inliers is selected as the final result. The ground noise points in the final result are filtered out, and the outliers are filtered out using radius filtering to obtain the filtered point cloud. Then, Euclidean clustering is used to obtain the clustered point cloud clusters according to the set clustering sphere radius and clustering point cloud threshold. The height and width of the clustered point cloud clusters are restricted according to the cone barrel features, and obstacles outside the cone barrel are filtered out. The cone barrels that meet the feature information are screened out, and the extracted cone barrel point cloud and its position information are sent to the data fusion module.

[0043] The data fusion module receives the cone point cloud cluster and its three-dimensional bounding box obtained by the point cloud clustering module and the image bounding box obtained by the image processing module, and then projects the three-dimensional bounding box extreme value to the pixel plane to obtain a two-dimensional bounding box. Then, the Hungarian algorithm is used for matching calculation, and the distance intersection and union ratio is used as the allocation cost matrix to complete the matching of the image and point cloud detection results, thereby realizing the fusion association of the cone position information and the category information.

[0044] In summary, this embodiment uses the improved YOLOv7-tiny target detection algorithm and DeepSort target tracking module, combined with lidar point cloud processing, to achieve accurate recognition and tracking of track markers such as cones, which not only improves the real-time and accuracy of target detection, but also further enhances the robustness of environmental perception through data fusion technology.

[0045] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An unmanned driving control system, characterized in that: The system includes the following components: environmental perception layer, positioning construction layer, path planning layer, decision control layer and bottom execution layer; The environmental perception layer includes an image processing module, a point cloud processing module and a data fusion module, and is equipped with a high-precision laser radar and a camera. It uses an improved YOLOv7-tiny target detection algorithm and a DeepSort target tracking algorithm to achieve real-time detection and tracking of cone barrel targets. The positioning and construction layer adopts the SC-LeGO-LOAM laser SLAM algorithm to build a real-time map through the scanning data of the laser radar and perform precise positioning; The path planning layer discretizes the global map using Delaunay triangulation, generates the target path through a graph search tree algorithm, and uses a cubic B-spline curve for path fitting; The decision control layer uses a hybrid lateral control algorithm based on DDPG and a three-point curvature velocity planning method to control the front wheel steering angle and acceleration and deceleration of the vehicle according to the path information and vehicle status; The bottom execution layer receives control signals through the CAN network and precisely controls the steering system, the braking system and the driving system.

2. An unmanned driving control system according to claim 1, characterized in that: The environmental perception layer includes an image processing module, a point cloud processing module and a data fusion module. The image processing module obtains the 2D detection frame of the target by realizing target detection and tracking. The point cloud processing module realizes target clustering and extracts the target point cloud and 3D coordinate information. The data fusion module parses the coordinates, color and confidence information of the cone barrel in the current field of view by fusing vision and lidar data, and provides it to the path planning layer for use.

3. An unmanned driving control system according to claim 2, characterized in that: The image processing module in the environment perception layer uses an improved YOLOv7-tiny target detection algorithm to achieve target detection and tracking. Specifically, image data containing a target object is received and processed using a model. The model includes a feature extraction layer, a target classification layer, and a bounding box regression layer. In the target classification layer, the classification loss L is calculated using a cross entropy loss function. cls , in the bounding box regression layer, the bounding box regression loss L is calculated by the squared error loss function bbox , in order to optimize the model's ability to predict the target bounding box, the classification loss and the bounding box regression loss are weighted summed to obtain the total loss, which is calculated as: L total =αL cls +βL bbox , where α and β are hyperparameters, and the model is trained and optimized based on the total loss. Based on the real box and the predicted box, the DeepSort target tracking algorithm is used to calculate the CIoU value between each detection box and each predicted box, and the allocation cost matrix is ​​constructed. The calculation formula of CIoU is: Among them, IoU is the intersection over union ratio, b and b gt are the center points of the predicted box and the real box respectively, ρ 2 (b,b gt ) is the square of the Euclidean distance between the center points, c is the diagonal distance of the minimum closure area, v is the aspect ratio consistency measure, α is the trade-off parameter, and the Hungarian algorithm is used to match the detection box with the prediction box based on the above allocation cost matrix, and a cascade matching mechanism is introduced to improve the matching reliability. Finally, the Kalman filter is updated with the successfully associated detection box, and the updated bounding box is passed to the next frame as the tracking box. By combining the recognition result with the target tracking result, the recognition of the cone bucket color can be visualized based on the tracking count. At the same time, when the target detector fails to detect the target, the position of the cone bucket target can still be predicted.

4. The unmanned driving control system according to claim 2, characterized in that: The point cloud processing module in the environmental perception layer filters out the point clouds outside the rectangular area where the cones are distributed according to the situation of the stadium, and uses the RANSAC algorithm to select the point cloud with the largest number of random fittings in the iterative process according to the set threshold as the extracted ground point cloud, filter out the ground noise points, and use radius filtering to filter out the outliers to obtain the filtered point cloud, and then use Euclidean clustering to obtain the clustered point cloud cluster according to the set clustering sphere radius and clustering point cloud threshold, and limit the height and width of the clustered point cloud cluster according to the cone characteristics, filter out obstacles outside the cone, screen out the cones that meet the feature information, and send the extracted cone point cloud and its position information to the data fusion module.

5. The unmanned driving control system according to claim 2, characterized in that: The environmental perception layer data fusion module receives the cone bucket point cloud cluster and its three-dimensional bounding box obtained by the point cloud clustering module and the image bounding box obtained by the image processing module, projects the three-dimensional bounding box extreme value to the pixel plane to obtain a two-dimensional bounding box, and then uses the Hungarian algorithm to perform matching calculations, and uses the intersection and union ratio of the distances between the three-dimensional bounding box and the two-dimensional bounding box as the allocation cost matrix to complete the matching of the image and point cloud detection results, thereby realizing the fusion association of the cone bucket position information and the category information.

6. The unmanned driving control system according to claim 1, characterized in that: The positioning and construction layer adopts the SC-LeGO-LOAM laser SLAM algorithm, builds a real-time map through the scanning data of the laser radar, and performs precise positioning. Specifically, the laser radar scans the surrounding environment and generates point cloud data, and extracts feature points from the point cloud data. The real-time map is built through the feature matching algorithm. The map and the current laser scanning data are used to perform precise positioning through the laser SLAM algorithm. The laser radar scanning data is set to S = {p1, p2, ..., pN}, where pi is a point in the scan with polar coordinates (ri, θi), and θ is divided into M intervals of equal width, and the width of each interval is For each interval j, calculate the maximum and minimum radius r of the points that fall within the interval max,j , r min,j , the Scan Context algorithm is used to process the scan data. The Scan Context histogram H is an M-dimensional vector, where H[j] = (r min,j ,r max,j ), loops are identified by comparing the similarity between the current scan and the historical scan. Suppose the two Scan Context histograms are H1 and H2, then the similarity D between them can be obtained by calculating the Euclidean distance between the corresponding intervals of the two histograms, that is, 7. The unmanned driving control system according to claim 1, characterized in that: The path planning layer uses the Delaunay triangulation algorithm to discretize the global map into a grid composed of points and lines. In the discretized grid, the graph search tree algorithm is used to generate the optimal path from the starting point to the end point based on the actual cost and estimated cost between nodes. The generated path points are used as control points and fitted using a cubic B-spline curve to obtain a smooth path with continuous curvature. The expression of the cubic B-spline curve is: Where P(t) is the position of any point on the path, B j,3 (t) is the cubic B-spline basis function, P i+j It is the control point.

8. The unmanned driving control system according to claim 1, characterized in that: The decision control layer receives the path information generated by the path planning layer and the current state information of the vehicle, and adopts a hybrid lateral control algorithm based on DDPG according to the current speed of the vehicle. The Stanley algorithm is used to calculate the front wheel angle in the low-speed range, and the MPC algorithm is used to calculate the front wheel angle in the high-speed range. The weights of the Stanley and MPC algorithms are dynamically adjusted by the DDPG algorithm in the transition range to achieve a smooth transition. By adopting a three-point curvature speed planning method, the speed of the vehicle is calculated and planned in sections according to the curvature change of the path, so that the vehicle can accelerate and decelerate smoothly during driving, and the calculated front wheel angle and speed are used as control instructions, which are sent to the underlying actuator through the vehicle control system to achieve control of the front wheel angle and acceleration and deceleration of the vehicle.

9. The unmanned driving control system according to claim 1, characterized in that: The bottom execution layer adopts the PID control algorithm, and according to the error between the expected angle and the actual angle and the rate of change thereof, calculates the output angle and controls the steering actuator by adjusting the proportional, integral and differential gains. The fuzzy PID control algorithm is adopted, and the actual speed, expected speed, speed error and its rate of change are input, and the PID parameters are dynamically adjusted through fuzzy reasoning to achieve precise control of the vehicle braking system.

Images