An intelligent driving obstacle avoidance analysis system

Abstract

The application relates to the technical field of intelligent driving and discloses an obstacle avoidance analysis system for intelligent driving. The application comprises data acquisition, road feature extraction, lane type determination, virtual marking line generation, obstacle detection and trajectory prediction, dynamic obstacle avoidance decision, and system verification and optimization modules. Data are collected by a laser radar, a camera and the like, road width and curvature are extracted by fusing point clouds and images, lane types are determined according to the width, virtual marking lines are generated on roads without marking lines, B-spline curves are used for fitting when the curvature is large, obstacles are detected by fusing point clouds and images, trajectories are predicted by Kalman filtering or a social force model, safe trajectories are generated by an OccupancyGrid map, an A* algorithm and MPC in combination with virtual marking lines and risk assessment, the system is optimized through simulation and real road testing, trajectory planning deviation on roads without marking lines is solved, and obstacle avoidance precision and road condition adaptability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, and in particular to an obstacle avoidance analysis system for intelligent driving. Background Technology

[0002] The development of intelligent driving technology has extended from structured roads to unstructured roads. Currently, mainstream autonomous driving systems rely on visual or radar features such as lane lines and traffic signs for environmental perception and path planning. For example, they use cameras to identify lane markings and use lidar point clouds to fit road boundaries, thereby achieving functions such as lane keeping and obstacle avoidance.

[0003] However, in unstructured scenarios such as rural dirt roads and construction zones, traditional technical solutions face severe challenges. Current mainstream systems use cameras for visual semantic segmentation to identify lane boundaries or use LiDAR point cloud data to fit road contours, but these methods have significant shortcomings in environments where road markings are missing or blurred. Specifically, visual systems easily misjudge non-road features such as roadside vegetation and temporary soil piles as road extensions; while LiDAR can acquire 3D point clouds, it is difficult to accurately distinguish between drivable areas and obstacle areas due to the lack of road marking features as a reference. These problems collectively lead to the system's susceptibility to missed or false detection of obstacles and trajectory planning deviations on unmarked roads, seriously affecting driving safety and comfort. Existing technologies have not yet effectively solved the problem of coordinating virtual lane construction and dynamic obstacle avoidance decision-making in unstructured road environments, especially in terms of balancing trajectory smoothness and real-time obstacle avoidance requirements on roads with complex curvature. Therefore, existing technologies urgently need improvement to address these issues. Summary of the Invention

[0004] The technical problem to be solved by this invention is the shortcomings of the prior art. To address this, we propose an obstacle avoidance analysis system for intelligent driving.

[0005] To achieve the above objectives, this application adopts the following technical solution: an obstacle avoidance analysis system for intelligent driving, comprising: Data acquisition module: used to collect road environment data through lidar, camera, millimeter-wave radar and IMU, and to achieve hardware and software synchronization of multiple sensors based on GPS time synchronization and ROS timestamp interpolation algorithm; Road feature extraction module: used to extract road width W and curvature k from the collected data. The road width is obtained by variance fusion of point cloud method and image method, and the curvature is calculated by fitting a quadratic curve by sliding window least squares method. Lane type determination module: used to determine the lane type based on the road width W, and to correct the determination result by combining historical driving data and vehicle distribution information of adjacent lanes; Virtual lane marking generation module: Used to generate dynamically adapted virtual lane markings based on lane type and curvature features on roads without lane markings, including: The number and location of lane markings are determined based on lane type; Align the point cloud with the image coordinate system using the extrinsic parameter matrix; A piecewise sliding window is used to fit the boundary point cloud; When the curvature k is too large, B-spline curves are used to generate smooth lines. Obstacle detection and trajectory prediction module: used to detect obstacles by fusing point cloud and image data, and predict trajectories based on Kalman filtering or social force models; Dynamic obstacle avoidance decision module: used to generate a safe trajectory based on virtual markings and obstacle risk assessment results, using OccupancyGrid map, A* algorithm and MPC; System verification and optimization module: used to verify performance through CARLA simulation and real-world testing, and to optimize parameters using a genetic algorithm.

[0006] Preferably, in the road feature extraction module, the variance fusion formula for the road width W is: ,in To remove the bottom point cloud, retain the road boundary point cloud, separate the road boundary point cloud clusters, and extract the road edge contour to distinguish between road and non-road areas, the lateral coordinate difference of the road boundary point cloud is then calculated using the point cloud method. The pixel width of the road area after perspective transformation is calculated using image methods, and the physical width is converted using camera calibration parameters. To measure variance.

[0007] Preferably, the lane type determination module executes the following determination logic: like It was determined to be a single lane; like It was determined to be a two-lane road; like It was determined to be a four-lane road; like It was determined to be a multi-lane road.

[0008] Preferably, in the virtual marking generation module, the marking position generation rules include: For a two-lane road: extract the point clouds of the left and right boundaries of the road, and then extract the point clouds along the center line of the road. Generate a single virtual line at the specified location; For a four-lane road: Generate three virtual lane markings, located at the following positions: , , ,in The location simulates the double yellow lines.

[0009] Preferably, the virtual marking generation module performs the following: Coordinate transformation: Align the boundary coordinates of the LiDAR point cloud with the camera image coordinates using an extrinsic parameter matrix; Piecewise fitting: A quadratic curve is fitted piecewise over a sliding window of length L to the lane boundary point cloud. Using the least squares error function Solve for the coefficients a, b, and c of the curve using the curvature formula: Calculate the curvature of the road at a certain point, where x is the x-coordinate of the center point of the window.

[0010] Preferably, in the virtual marking generation module, the B-spline curve is generated in the following way: ,in: This represents the coordinate value of the B-spline curve at parameter u; Denotes the i-th p-th B-spline basis function; Represents the two-dimensional control points of the B-spline curve; The parameter u takes values ​​that cover the entire length of the road, and continuous road markings are generated by traversing u.

[0011] Preferably, the B-spline basis functions are generated using a recursive formula: 0th-order basis functions: ; p-th basis functions: .

[0012] Preferably, the obstacle avoidance strategy executed by the dynamic obstacle avoidance decision module is as follows: when the collision time of the obstacle in this lane is ≤3 seconds and the adjacent lane is safe, the DWA dynamic window method is used to perform lane change; otherwise, braking is triggered.

[0013] Preferably, the path planning process of the dynamic obstacle avoidance decision module includes: Construct an OccupancyGrid grid map and expand the obstacle region; The A* algorithm is used to search for a global path based on virtual markings. The MPC model is used to predict and optimize the trajectory.

[0014] Preferably, the system verification and optimization module performs the following: Testing virtual lane marking errors and obstacle avoidance success rates on the CARLA platform; The MPC weight matrix is ​​optimized using a genetic algorithm.

[0015] The technical effects and advantages of this invention are as follows: In this invention, a virtual lane centerline is generated by equally dividing the road width and fitting the curvature, replacing the traditional physical markings as a reference for the driving trajectory. This allows the intelligent driving system to clearly define the boundaries of the drivable area on unstructured roads such as rural dirt roads without markings, roads with worn or missing markings, and construction sections. It overcomes the limitations of existing technologies that rely on structured features such as lane lines, and enables autonomous driving path planning in unmarked environments. It also solves the problem of trajectory planning deviation caused by the lack of lane references in unstructured roads, and has the advantages of improving obstacle detection accuracy, optimizing driving trajectory smoothness, and enhancing adaptability to complex road conditions. Attached Figure Description

[0016] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts: Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0017] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0018] Reference Figure 1 As shown, the present invention provides a technical solution: an obstacle avoidance analysis system for intelligent driving, comprising: Data acquisition module: Real-time acquisition of multi-dimensional data of the road environment, including road geometry information, dynamic obstacle data and vehicle status, providing raw data support for subsequent analysis; real-time acquisition of road information through LiDAR, camera and millimeter-wave radar sensors, acquisition of vehicle acceleration, angular velocity and attitude angle information through IMU inertial measurement unit, and sensor hardware synchronization through GPS timing module, and software synchronization through ROS timestamp interpolation algorithm; The system utilizes LiDAR, cameras, and millimeter-wave radar sensors. Specifically, it can be implemented using a combination of Velodyne 16-line LiDAR, a binocular stereo camera, and Continental ARS408 millimeter-wave radar. By leveraging the complementary characteristics of different sensors, it covers the visible light, infrared, and microwave frequency bands, solving the problem of blind spots in rainy, foggy, or low-light conditions caused by single sensors. The IMU (Inertial Measurement Unit) is a vehicle motion state detection device, which can be implemented using the MPU-6050 six-axis sensor module. It captures vehicle dynamic parameters in real time using a three-axis accelerometer and a three-axis gyroscope, providing a motion compensation benchmark for road feature extraction. The GPS timing module and ROS timestamp interpolation algorithm are a spatiotemporal synchronization device, which can be implemented using the Ublox NEO-M8N module in conjunction with the ROS system's timestamp alignment function. It triggers sensor synchronous acquisition through hardware PPS pulse signals, and then corrects the time deviation of the software layer through a cubic spline interpolation algorithm, solving the fusion error problem caused by spatiotemporal misalignment of multi-source data.

[0019] Specifically, by combining LiDAR, cameras, and millimeter-wave radar, road information is collected in multiple dimensions. LiDAR generates high-precision point cloud data for road boundary detection, cameras capture RGB images for semantic segmentation, and millimeter-wave radar continuously tracks dynamic obstacles. The IMU (Inertial Measurement Unit) outputs vehicle triaxial acceleration and angular velocity data at a fixed frequency, which is fused with the pose information output by the visual SLAM algorithm using Kalman filtering to eliminate point cloud distortion caused by vehicle bumps. The GPS timing module generates a 1Hz PPS synchronization pulse signal to trigger data acquisition from each sensor under a unified time reference. The ROS system records the timestamps of each sensor data packet and uses interpolation algorithms to compensate for minor time deviations caused by communication delays, ensuring the consistency of multi-source data in the spatiotemporal dimensions.

[0020] The road feature extraction module extracts key geometric features such as road width, boundary contour, and curvature from the collected data, providing a quantitative basis for lane type determination and virtual marking generation. It removes the bottom point cloud, retains the road boundary point cloud, separates road boundary point cloud clusters, and extracts the road edge contour to distinguish between road and non-road areas. Subsequently, it calculates the lateral coordinate difference of the road boundary point cloud using a point cloud method. The pixel width of the road area after perspective transformation is calculated using image-based methods, and then converted into physical width using camera calibration parameters. The variances measured by the two methods are fused. To obtain a more accurate road width W, To measure variance, curvature is calculated by fitting a quadratic curve to the road boundary point cloud or image boundary coordinate sequence using the sliding window least squares method. Least squares error function Solve for the coefficients a, b, and c of the curve using the curvature formula: Calculate the curvature of the road at a certain point, where x is the x-coordinate of the center point of the window; Specifically, the road feature extraction module first removes ground interference points through point cloud segmentation, retaining the point cloud clusters at the road boundary. The point cloud method calculates the initial value of the road width by statistically analyzing the extreme differences of the lateral coordinates of the boundary points, while the image method measures the pixel width and converts it into the physical width after eliminating distortion through inverse perspective transformation. The measurement results of the two methods are fused using variance weighting. For example, the weight of the point cloud is reduced when the point cloud density is low, and the weight of the image method is reduced when the lighting conditions are poor. The final output is the fused road width. For curvature calculation, a sliding window is used to divide the road boundary points into multiple local intervals. Within each window, a quadratic curve is fitted using the least squares method. The curve coefficients are extracted and then substituted into the curvature formula to calculate the curvature value of the center point of the current window, thereby constructing the curvature distribution of the entire road.

[0021] Lane type determination module: Determines the lane type based on the road width, including single lane, two lane and four lane, providing a decision basis for virtual marking generation; like It was determined to be a single lane; like It was determined to be a two-lane road; like It was determined to be a four-lane road; like It was determined to be a multi-lane road; The lane type determination result is corrected by combining historical driving data and vehicle distribution information in adjacent lanes; Virtual lane marking generation module: Generates dynamically adapted virtual lane centerlines on roads without lane markings based on lane type and curvature characteristics; For a two-lane road: extract the point clouds of the left and right boundaries of the road, and then extract the point clouds along the center line of the road. Generate a single virtual line at the specified location; For a four-lane road: Generate three virtual lane markings, located at the following positions: , , , The location simulates the double yellow lines; Specifically, in unmarked road scenarios, the system first obtains accurate road width measurements using a multi-sensor fusion algorithm. Then, it matches the real-time measurements with a preset width threshold range. For example, when the measured width is 5 meters, the current road is determined to be a two-lane road. When the measured width is close to the threshold boundary, such as when the measured width is 7.9 meters, the system retrieves historical driving trajectory data for that road segment to analyze whether the actual driving path of the vehicle exhibits two-lane characteristics, thereby avoiding misjudgment of lane type due to single measurement errors. When the determination result is multi-lane, the system further combines the typical driving trajectory distribution patterns of vehicles in that area from historical data, such as whether there are characteristics of multiple vehicles driving in parallel, to perform a secondary verification of the preliminary determination result.

[0022] Perform coordinate transformation: Align the point cloud boundary coordinates with the camera image coordinates using an extrinsic parameter matrix to ensure that the position of the marking line is consistent in both the visual and lidar coordinate systems; Segmented sliding window fitting: For each lane boundary point cloud or image boundary coordinate sequence, a sliding window of length L is used to fit a quadratic curve segment by segment. Curvature is calculated using the curvature calculation formula; When curvature At that time, the virtual marking line is a straight line; When curvature When using B-spline curve fitting: A continuous curve is generated by iterating through the parameter u, where the range of u values ​​corresponds to the road length, ensuring that the markings cover the entire road segment; and the control points of the B-spline curve are adjusted to ensure a smooth transition of the markings. This represents the coordinate value of the B-spline curve at parameter u. This formula is used to calculate the points corresponding to different u values ​​on the curve, thus determining the shape of the B-spline curve. In virtual road marking generation, it is used to fit virtual road markings with different curvatures, allowing the markings to smoothly change according to the road's curvature. By changing the value of u, various points on the B-spline curve can be traversed, thereby generating a complete curve. In practical applications of virtual road marking generation, by taking values ​​for u within a certain range, the corresponding coordinate points are calculated, and these points are connected to form virtual markings that conform to the road's curvature. Let represent the i-th p-th B-spline basis function, which is a function of the parameter u and determines each control point. The basis functions contribute to the curve at u and exhibit local support, meaning that for a specific value of u, only a few basis functions are available. The value is not zero, which means that when modifying individual control points, only the local shape of the curve will be affected, ensuring the smoothness and controllability of the curve; The control points representing the B-spline curve are points in three-dimensional space. In a two-dimensional road marking scenario, they are simplified to two-dimensional coordinate points. There are a total of n+1 control points. The position of these control points determines the approximate shape of the B-spline curve. By adjusting the coordinates of the control points, the shape of the curve can be flexibly changed to adapt to different curvatures of the road, making the generated virtual markings more in line with actual driving needs. This represents a summation operation, which sums the values ​​of each control point. Its corresponding basis functions The results of multiplication are summed to obtain the coordinates of the B-spline curve at parameter u; The recursive formula for the B-spline basis function is: and , This is the initial definition of the B-spline basis function, i.e., a 0th-order B-spline basis function, which determines the support interval of the basis function when the parameter u is located in the nodal interval. hour, It is 1 in the interval and 0 in other intervals. This is a recursive formula for B-spline basis functions, used to calculate higher-order basis functions from lower-order basis functions. Here, p represents the degree of the B-spline curve. As p increases, the smoothness and complexity of the curve also increase accordingly. The formula calculates higher-order basis functions from two lower-order basis functions. and By performing weighted combinations, we obtain basis functions of a higher order. The weights are respectively from and It is confirmed that these two weighting coefficients are related to the node values. This ensures the continuity and smoothness of the basis functions, thereby guaranteeing that the generated B-spline curves can smoothly transition when fitting the virtual calibration line, meeting the requirements of vehicle driving dynamics. Specifically, during the virtual lane marking generation process, when the road curvature reaches a set threshold, the coordinates of the control points of the B-spline curve are input into the basis function calculation module. The node sequence is dynamically divided according to the road boundary curvature distribution. The 0th-order basis function first generates an initial weight distribution based on the node interval to which the parameter variables belong. Then, the 0th-order basis functions of adjacent intervals are weighted and combined using a recursive formula to generate a first-order basis function. This process iterates layer by layer until the target order is reached. Finally, the basis function values ​​of each control point are weighted and summed to output continuous curve coordinates, achieving a smooth fit of the virtual marking. This calculation method enables the curve shape to adaptively adjust according to changes in road curvature, while avoiding the Runge phenomenon present in traditional polynomial fitting.

[0023] Obstacle detection and trajectory prediction module: Real-time detection of road obstacles, prediction of their trajectory, and assessment of collision risk. It detects the 3D position, size, and category of obstacles through point cloud and detects 2D coordinates by combining image data. Based on spatial position and feature similarity, it correlates the detection results. For vehicles, it uses Kalman filtering; for pedestrians, it uses a social force model to calculate dynamic risk. Specifically, the obstacle detection and trajectory prediction module acquires the three-dimensional coordinates and geometric features of obstacles through point cloud sensors, and performs semantic classification using two-dimensional image data captured by cameras. For vehicle-type obstacles, a Kalman filter is used to predict their future movement path based on historical trajectory data. For unstructured moving targets such as pedestrians, a social force model is used to simulate their movement trends under the influence of the environment. By fusing point cloud and image detection results, a joint estimation of the spatial position and motion state of obstacles is established. The time-of-flight collision (TTC) risk calculation module monitors the relative motion relationship between the vehicle and each obstacle in real time, and triggers the obstacle avoidance decision mechanism when the TTC value is lower than a preset threshold.

[0024] Dynamic obstacle avoidance decision module: Based on virtual markings, obstacle risk assessment results, and vehicle status, it dynamically generates a safe driving trajectory and outputs control commands; by constructing an OccupancyGrid grid map, it expands the obstacle area, models the drivable area, then searches for a preliminary path using the global A* algorithm, predicts and optimizes the trajectory using the MPC model, and provides obstacle avoidance strategies. Specifically, in unmarked road scenarios, the OccupancyGrid grid map converts the sensor-perceived obstacle locations into a computable grid model, and inflates the obstacle areas to increase safety margins. During the global path planning phase, the A* algorithm searches for the optimal path from the starting point to the destination in the grid map, generating a preliminary path including key turning points. In the local trajectory optimization phase, the MPC model predictive control algorithm performs rolling optimization on the vehicle's trajectory for several future steps, ensuring a smooth trajectory that conforms to vehicle dynamics constraints. When the obstacle risk assessment module detects that the collision time of an obstacle in the current lane is less than or equal to 3 seconds, the system prioritizes assessing the safety status of adjacent lanes. If the lane change conditions are met, the DWA dynamic window method is activated, generating multiple candidate trajectories in the speed-steering angle joint space, and selecting the optimal lane change path based on the trajectory score. If there is a risk in adjacent lanes, an emergency braking strategy is triggered.

[0025] If the obstacle is in the lane and the collision time is less than or equal to three seconds, and the adjacent lane is safe, then change lanes; otherwise, brake. When changing lanes, use the DWA dynamic window method to evaluate candidate paths. System verification and optimization module: Verify system performance through simulation and real-road testing, and optimize algorithm parameters and models; evaluate virtual marking errors, obstacle avoidance success rate, and comfort indicators by simulating different scenarios on the CARLA platform; collect data on rural roads and compare the deviation between actual trajectories and virtual markings; optimize the MPC weight matrix using a genetic algorithm.

[0026] Specifically, the system verification and optimization module's workflow is divided into two stages: simulation verification and real vehicle testing. In the simulation stage, complex scenarios such as unmarked roads and dynamic obstacles are constructed using the CARLA platform. The lateral deviation between the path output by the virtual marking generation module and the preset road boundary is recorded. At the same time, the success rate of the obstacle avoidance decision module under different collision risk thresholds is statistically analyzed. In the real vehicle testing stage, trajectory data is collected in real time when the vehicle is driving on rural roads. The average offset is calculated after spatially aligning the actual driving path with the virtual marking coordinates. This offset is used to correct the control parameters of the curvature fitting algorithm. For the optimization problem of the model predictive controller, a multi-objective function including trajectory tracking error and acceleration change rate is established. The weight matrix is ​​globally optimized using a genetic algorithm, so that the vehicle can balance safety and comfort during obstacle avoidance.

[0027] In a two-lane scenario, the actual curvature of the virtual markings is calculated by merging the curvature of the left road boundary and the curvature of the centerline in a weighted manner. For example, the curvature of the left boundary is weighted at 0.7 and the curvature of the centerline is weighted at 0.3, so that the virtual markings not only conform to the shape of the actual road boundary, but also take into account the need for the vehicle to stay in the center. After generating the B-spline curve, the autonomous driving system calculates its jerk value to judge the smoothness of the trajectory. If the jerk value exceeds the preset threshold, the control point position is readjusted to reduce acceleration abrupt changes. For example, the gradient descent method is used to iteratively optimize the control point coordinates until the jerk value meets the comfort requirements.

[0028] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An obstacle avoidance analysis system for intelligent driving, characterized in that, include: Data acquisition module: used to collect road environment data through lidar, camera, millimeter-wave radar and IMU, and to achieve hardware and software synchronization of multiple sensors based on GPS time synchronization and ROS timestamp interpolation algorithm; Road feature extraction module: used to extract road width W and curvature k from the collected data. The road width is obtained by variance fusion of point cloud method and image method, and the curvature is calculated by fitting a quadratic curve by sliding window least squares method. Lane type determination module: used to determine the lane type based on the road width W, and to correct the determination result by combining historical driving data and vehicle distribution information of adjacent lanes; Virtual lane marking generation module: Used to generate dynamically adapted virtual lane markings based on lane type and curvature features on roads without lane markings, including: The number and location of lane markings are determined based on lane type; Align the point cloud with the image coordinate system using the extrinsic parameter matrix; A piecewise sliding window is used to fit the boundary point cloud; When the curvature k reaches the set threshold, a smooth line is generated using a B-spline curve. In a two-lane scenario, the virtual road marking generation module weights the curvature of the left road boundary with the curvature of the center line to determine the actual curvature of the virtual road marking. The node sequence of the B-spline curve is dynamically divided according to the curvature distribution of the road boundary, and each control point is multiplied by the corresponding B-spline basis function and then accumulated to generate a continuous virtual marking line. After generating the B-spline curve, the virtual marking generation module calculates the jerk value of the trajectory corresponding to the B-spline curve. When the jerk value exceeds a preset threshold, the gradient descent method is used to iteratively adjust the coordinates of the control points of the B-spline curve until the jerk value does not exceed the preset threshold. Obstacle detection and trajectory prediction module: used to detect obstacles by fusing point cloud and image data, and predict trajectories based on Kalman filtering or social force models; Dynamic obstacle avoidance decision module: used to generate a safe trajectory based on virtual markings and obstacle risk assessment results, using OccupancyGrid map, A* algorithm and MPC; System verification and optimization module: used to verify performance through CARLA simulation and real-world testing, and to optimize parameters using a genetic algorithm; The system verification and optimization module is also used to spatially align the actual driving path and virtual marking coordinates collected in the real road test, calculate the average offset between the actual driving path and the virtual marking, and correct the control parameters of the curvature fitting algorithm based on the average offset.

2. The obstacle avoidance analysis system for intelligent driving according to claim 1, characterized in that: In the road feature extraction module, the variance fusion formula for road width W is: ,in To remove the bottom point cloud, retain the road boundary point cloud, separate the road boundary point cloud clusters, and extract the road edge contour to distinguish between road and non-road areas, the lateral coordinate difference of the road boundary point cloud is then calculated using the point cloud method. The pixel width of the road area after perspective transformation is calculated using image methods, and the physical width is converted using camera calibration parameters. To measure variance.

3. The obstacle avoidance analysis system for intelligent driving according to claim 1, characterized in that: The lane type determination module executes the following determination logic: like It was determined to be a single lane; like It was determined to be a two-lane road; like It was determined to be a four-lane road; like It was determined to be a multi-lane road.

4. The obstacle avoidance analysis system for intelligent driving according to claim 1, characterized in that: The virtual marking generation module includes the following rules for generating marking positions: For a two-lane road: extract the point clouds of the left and right boundaries of the road, and then extract the point clouds along the center line of the road. Generate a single virtual line at the specified location; For a four-lane road: Generate three virtual lane markings, located at the following positions: , , ,in The location simulates the double yellow lines.

5. The obstacle avoidance analysis system for intelligent driving according to claim 1, characterized in that: The virtual marking generation module performs the following: Coordinate transformation: Align the boundary coordinates of the LiDAR point cloud with the camera image coordinates using an extrinsic parameter matrix; Piecewise fitting: A quadratic curve is fitted piecewise over a sliding window of length L to the lane boundary point cloud. Using the least squares error function Solve for the coefficients a, b, and c of the curve using the curvature formula: Calculate the curvature of the road at a certain point, where x is the x-coordinate of the center point of the window.

6. An obstacle avoidance analysis system for intelligent driving according to claim 1 or 5, characterized in that: The virtual marking generation module generates B-spline curves in the following way: ,in: This represents the coordinate value of the B-spline curve at parameter u; Denotes the i-th p-th B-spline basis function; Represents the two-dimensional control points of the B-spline curve; The parameter u takes values ​​that cover the entire length of the road, and continuous road markings are generated by traversing u.

7. The obstacle avoidance analysis system for intelligent driving according to claim 6, characterized in that: The B-spline basis functions are generated using a recursive formula: 0th-order basis functions: ; p-th basis functions: .

8. The obstacle avoidance analysis system for intelligent driving according to claim 1, characterized in that: The obstacle avoidance decision module executes the following obstacle avoidance strategy: when the collision time of the obstacle in this lane is ≤3 seconds and the adjacent lane is safe, the DWA dynamic window method is used to execute the lane change; otherwise, braking is triggered.

9. The obstacle avoidance analysis system for intelligent driving according to claim 1, characterized in that: The path planning process of the dynamic obstacle avoidance decision module includes: Construct an OccupancyGrid grid map and expand the obstacle region; The A* algorithm is used to search for a global path based on virtual markings. The MPC model is used to predict and optimize the trajectory.

10. The obstacle avoidance analysis system for intelligent driving according to claim 1, characterized in that: The system verification and optimization module performs the following: Testing virtual lane marking errors and obstacle avoidance success rates on the CARLA platform; The MPC weight matrix is ​​optimized using a genetic algorithm.

Images