Automatic homeward voyage method for multi-sensor data fusion of unmanned vehicle

Abstract

The invention relates to the technical field of automatic driving, in particular to an automatic homeward voyage method for multi-sensor data fusion of an unmanned vehicle, which comprises the following steps: acquiring laser radar obstacle points, image edges and attitude data, completing synchronous registration, identifying space mapping, extracting track and boundary trends, screening forward sections, and judging track offset and trafficability. Screening an error band boundary, adjusting a steering instruction, analyzing path overlapping change, and outputting a path keeping label. According to the method, through time synchronization and space registration of various sensor data, high-precision fusion of the data under a unified space-time framework is realized, the reliability of environment perception is improved, accurate mapping of the data in a unified reference coordinate system is ensured, the reliability of path planning and decision making is enhanced, and the optimal advancing direction is determined by combining a homing point; the accuracy of return route planning is improved, the traffic feasibility is dynamically evaluated, the effective error band boundary is screened to improve the safety, and the reliability of the return route is ensured through track stability analysis.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an automatic return method for an unmanned vehicle by fusing multi-sensor data. Background Art

[0002] The field of autonomous driving encompasses technologies that enable vehicles to drive autonomously without human intervention through a variety of sensing methods and intelligent decision-making systems. Core elements of this technology include key aspects such as environmental perception, path planning, motion control, and decision execution. Autonomous driving systems rely on multiple sensors, including cameras, radar, lidar, GPS, and inertial navigation, to collaboratively collect information about the surrounding environment and their own status. This information is then integrated and processed to generate an accurate environmental model, which is then used to dynamically adjust and control the driving path. The systematic nature of this technology is reflected in the close coordination of multi-sensor fusion, collaborative optimization of perception and decision-making algorithms, automatic control execution mechanisms, and on-board communications and remote management platforms, enabling the stable operation of autonomous driving in complex traffic environments.

[0003] Among them, the automatic return method of multi-sensor data fusion for unmanned vehicles refers to the problem that the unmanned vehicle needs to return to the original starting point due to signal loss, path deviation or mission interruption during the execution of the mission. It uses the position information, motion information and environmental data collected by multiple heterogeneous sensors to determine the current position and return path of the unmanned vehicle through fusion weighted processing, and combines the prior path information with the original trajectory data for dynamic comparison and analysis to generate return trajectory instructions. This method mainly relies on the joint fusion of inertial navigation data and GPS position information, the recursive planning of the return path nodes, and the real-time correction of position errors during the return process to complete the whole process guidance.

[0004] While existing technologies in the autonomous driving field achieve environmental perception and path planning through multiple sensors, their sensor data fusion methods rely on simple weighted superposition or static configuration, lacking refined processing for time synchronization and spatial registration. This results in positional errors and data offsets in highly dynamic scenarios. Return-to-home path planning is based on the original path or static planning, failing to incorporate real-time obstacle distribution and visual boundary information, resulting in insufficient flexibility and dynamic adjustment capabilities for path planning. The fusion of inertial navigation data and GPS position information often relies on fixed weights and fails to dynamically adjust based on the actual scenario, resulting in reduced positioning accuracy in complex environments. In existing automatic return-to-home methods, steering control adjustments also rely primarily on fixed rules, lacking adaptive correction mechanisms based on vision and sensor feedback. This makes it impossible to effectively respond to dynamic environmental changes and path deviations. Trajectory control often relies on fixed thresholds or simple convergence judgments for stability, making it difficult to adapt to complex path changes and ensuring the stability and safety of the return path. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and propose an automatic return method for unmanned vehicles based on multi-sensor data fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an automatic return method for an unmanned vehicle based on multi-sensor data fusion, comprising the following steps: S1: Obstacle point information is acquired through lidar, image edge contours are acquired through visual sensors, and attitude data is acquired through inertial navigation units. The three types of data are synchronized and registered in time and space, and the state of the target area associated with the return path is extracted. The spatial mapping model between sensor devices is identified and the reference coordinate system is unified to generate the return state identification matrix. S2: Based on the return state identification matrix, the positioning trajectory, obstacle distribution and image boundary trend are extracted, the forward angle segment of the passable space is screened, and the optimal travel direction is selected in combination with the target homing point to generate a passable return path interval; S3: Based on the traversable return path interval, by analyzing the lateral distance trend between the vehicle trajectory line and the interval centerline, combining the attitude angle change to determine the trajectory offset and amplitude, and combining the lateral spatial characteristics of the LiDAR to determine the feasibility of passage, screen the effective error band boundary, and generate the return tolerance control limit; S4: calling the return tolerance control limit, comparing the steering control input with the current attitude deflection direction, and adjusting the steering angle direction and amplitude if there is a deviation, and correcting the control instruction according to the visual sensor feedback to obtain the steering input adjustment instruction set.

[0007] As a further solution of the present invention, the return status identification matrix includes the target area position, attitude information, obstacle characteristics, and environmental boundary information; the passable return path interval includes the passable area, obstacle distribution, travel direction, and boundary contour; the return tolerance control limit includes the lateral distance threshold, attitude angle threshold, lateral space tolerance, and error band width; and the steering input adjustment instruction set includes the steering angle, steering direction, and steering correction instruction.

[0008] As a further solution of the present invention, the steps for obtaining the return state identification matrix are specifically as follows: S111: Obstacle point information is acquired through the lidar, image edge contours are acquired through the visual sensor, and attitude data is acquired through the inertial navigation unit. The data are jointly interpolated using timestamps to extract the spatial pose offset and sensor imaging time series alignment error of each time slice. The position information of the three types of data is used for coordinate normalization processing to generate a time series registration coordinate set. S112: Based on the overlapping area of ​​the obstacle point position and the image edge contour in the time-series registration coordinate set, combined with the attitude angular velocity component, extract the stable area center offset trajectory and the adjacent frame position transformation angle, identify the return path, and generate the return channel posture parameter group; S113: According to the posture information of each path segment in the return channel posture parameter group, the sensor configuration mapping relationship in the original coordinate system is analyzed, the path points are converted to a unified reference system, the channel is verified for distance and angle based on the path continuity and posture change, the dynamic trajectory segment and the feasible posture range are integrated, and a return state identification matrix is ​​generated.

[0009] As a further solution of the present invention, the steps for obtaining the passable return path interval are specifically as follows: S211: Based on the return state identification matrix, the output data of the positioning sensor, the lidar, and the visual sensor are called to extract the positioning trajectory, obstacle distribution, and environmental boundary trend. The obstacle positions are identified according to the obstacle distribution. The path segments affected by the obstacles in the positioning trajectory are detected, and the path segments overlapping with the obstacles are eliminated to generate an obstacle-free passage trajectory. S212: Calling the barrier-free passage trajectory, combining it with the boundary detection data of the visual sensor, identifying the forward angle segment of each trajectory segment, extracting the angle information, calculating the passage space optimization value of the segment, and screening out the optimal passage angle segment; S213: Call the optimal pass angle section, combine it with the target homing point, calculate the angle deviation between the target direction and the section direction, select the optimal path segment according to the deviation, and generate a passable return path interval.

[0010] As a further solution of the present invention, the step of obtaining the return tolerance control limit is specifically as follows: S311: Extracting the lateral distance between the vehicle trajectory and the centerline of the navigable return path interval, identifying the lateral distance value and time interval between the current frame and the previous frame, analyzing the lateral distance change trend, determining the lateral change direction and numerical intensity, and obtaining a lateral offset trend value; S312: Call the lateral offset trend value, combine it with the attitude angle change, the obstacle boundary distance and the trajectory boundary distance extracted from the laser radar lateral space feature, perform a joint operation, calculate the feasibility quantification value of the return path, judge the feasibility of the trajectory based on the comparison result with the lateral space threshold, and screen the error band boundary of the passable interval to generate the return tolerance control limit.

[0011] As a further solution of the present invention, the steps of acquiring the steering input adjustment instruction set are specifically as follows: S411: Invoking the return-to-home tolerance control limit, extracting the steering angle input and the yaw angle calculated by the fusion of the inertial navigation and the vehicle attitude sensor, performing sign consistency and angle difference determination, and if the direction does not match or the angle difference exceeds a threshold, determining the return-to-home control deviation and obtaining the attitude control deviation status; S412: Based on the attitude control deviation state, combined with the road edge offset and steering feedback angle detected by the visual sensor, the steering input amplitude is dynamically adjusted through multi-sensor data fusion, and the return trajectory steering correction instruction is calculated. The instruction is fed back to the vehicle control, superimposed on the original control amount, and updated in real time to obtain a steering input adjustment instruction set.

[0012] As a further embodiment of the present invention, the method further comprises step S5: S5: Based on the steering input adjustment instruction set, collect the vehicle trajectory output of continuous cycles and the change in the length of the overlapping area of ​​the path line in the visual image, analyze whether the change trend is in a convergence state, determine the stability period after the control is executed, and output the return path maintenance status label; The return path maintenance status label includes track overlap rate, path stability indicator, and visual path consistency.

[0013] As a further solution of the present invention, the step of obtaining the return path holding status tag is specifically as follows: S511: Based on the steering input adjustment instruction set, collecting the overlapping area between the vehicle trajectory output by the positioning unit and the path detected by the vision unit, extracting the lateral offset and overlapping length of the trajectory, analyzing the variation amplitude of the overlapping length within the period, and generating a periodic overlapping amplitude sequence; S512: Calling the periodic overlapping amplitude sequence, extracting the overlapping amplitude difference sequence within consecutive periods, marking the stable trend according to the polarity of the difference change, and combining the set visual trajectory stability threshold to determine whether the overlapping change converges to a single trend, thereby obtaining the visual trajectory fusion stable trend value; S513: Based on the visual trajectory fusion stability trend value, the lateral difference between the inertial navigation pose solution and the visual positioning coordinates within the stable period is identified, and combined with the ratio of the length of the path overlap area, it is determined whether the unmanned vehicle maintains stable path tracking, and the return path maintenance status label is output.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, by acquiring multiple sensor data and achieving time synchronization and spatial registration, it is ensured that various types of data are effectively integrated in a unified spatiotemporal framework, thereby improving the accuracy and reliability of environmental perception. By extracting target area status information and constructing a spatial mapping model between sensor devices, accurate mapping of data in a unified reference coordinate system is achieved, providing a reliable basis for subsequent path planning and decision-making. By extracting positioning trajectories, obstacle distribution, and image boundary trends, and combining them with the target homing point to determine the optimal direction of travel, the accuracy and feasibility of return path planning are improved. Combining the lateral distance trend between the vehicle trajectory and the centerline of the interval, the attitude angle change, and the lateral spatial characteristics of the laser radar, the feasibility of passage can be dynamically evaluated, and the safety and fault tolerance of the return path are improved by screening the effective error band boundary. By dynamically adjusting the steering control input and combining it with the visual sensor feedback, adaptive steering input correction is achieved, thereby improving the accuracy of steering control. By analyzing the stability trends of the overlapping areas between the vehicle trajectory output and the path lines in the visual image, the stability and controllability of the return path are guaranteed. This comprehensive processing logic of synchronous fusion of multi-sensor data, path planning based on multi-dimensional information, dynamic fault-tolerant control and adaptive steering correction can significantly improve the path planning accuracy, traffic safety and control stability of the unmanned vehicle during the return process. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the main steps of the present invention; Figure 2 This is a flow chart for obtaining the return state identification matrix in the present invention; Figure 3 This is a flowchart for obtaining a navigable return path interval in the present invention; Figure 4 This is a flow chart for obtaining the return tolerance control limit in the present invention; Figure 5 This is a flow chart for obtaining a steering input adjustment instruction set in the present invention; Figure 6 This is a flow chart for obtaining the return path maintenance status tag in the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0017] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0018] Example 1 See also Figure 1 The present invention provides a technical solution: an automatic return method for an unmanned vehicle based on multi-sensor data fusion, comprising the following steps: S1: Obstacle point information is acquired through lidar, image edge contours are acquired through visual sensors, and attitude data is acquired through inertial navigation units. The three types of data are synchronized in time and spatially registered to extract the target area status information associated with the return path. The spatial mapping model between sensor devices is identified and the reference coordinate system is unified to generate the return status identification matrix. S2: Based on the return state recognition matrix, the positioning trajectory, obstacle distribution, and image boundary trend are extracted to screen the forward angle segment of the passable space. The optimal travel direction is selected in combination with the target homing point to generate a passable return path interval. S3: Based on the traversable return path interval, the lateral distance trend between the vehicle trajectory and the interval centerline is analyzed. The trajectory offset and amplitude are determined based on the attitude angle change. The feasibility of passage is determined based on the lateral spatial characteristics of the LiDAR. The effective error band boundary is screened to generate the return tolerance control limit. S4: Call the return tolerance control limit and compare the steering control input with the current attitude deflection direction. If there is a deviation, adjust the steering angle direction and amplitude, and correct the control command based on the visual sensor feedback to obtain the steering input adjustment command set; S5: Based on the steering input adjustment instruction set, collect the continuous cycle of vehicle trajectory output and the change in the length of the overlapping area of ​​the path line in the visual image, analyze whether the change trend is in a convergence state, determine the stability period after control execution, and output the return path maintenance status label.

[0019] The return status identification matrix includes the target area location, attitude information, obstacle characteristics, and environmental boundary information. The navigable return path interval includes the navigable area, obstacle distribution, travel direction, and boundary contour. The return tolerance control limit includes the lateral distance threshold, attitude angle threshold, lateral space tolerance, and error band width. The steering input adjustment instruction set includes steering angle, steering direction, and steering correction instructions. The return path maintenance status label includes trajectory overlap rate, path stability indicator, and visual path consistency.

[0020] See also Figure 2 , the specific steps for obtaining the return status identification matrix are: S111: Obstacle point information is acquired through the lidar, image edge contours are acquired through the visual sensor, and attitude data is acquired through the inertial navigation unit. The data are jointly interpolated using timestamps to extract the spatial pose offset and sensor imaging time series alignment error of each time slice. The position information of the three types of data is used for coordinate normalization processing to generate a time series registration coordinate set. LiDAR scans the environment in real time to acquire information about obstacles around the vehicle, generating point cloud data. This point cloud data accurately describes the three-dimensional spatial position (e.g., distance and angle) of obstacles and can provide the location of obstacles around the vehicle. For example, LiDAR scans 10 times per second (with a sampling interval of 0.1 seconds), providing discrete spatial point coordinates. Vision sensors capture images with cameras and use edge detection algorithms (such as the Canny operator) to extract object outlines, resulting in image edge contours that reflect the two-dimensional shape characteristics of obstacles. Inertial navigation units (INS) provide real-time vehicle posture data, including angle, velocity, acceleration, and other information. All data is synchronized with a timestamp. To perform joint data interpolation, the data from different sensors must be aligned based on each sensor's timestamp. This can be achieved through methods such as linear interpolation. For example, LiDAR and vision sensors have different sampling frequencies, and this difference must be accounted for during data alignment. By analyzing the spatial pose differences between the sensor data, such as the distance difference between the obstacle position of the lidar and the image contour of the visual sensor, the offset of the spatial pose is calculated using the interpolation method. Then, by comparing each frame of data in the time series, the sensor imaging timing alignment error of each time slice is obtained and corrected. The position information of the lidar, visual sensor and inertial navigation unit is normalized (i.e. converted to the same coordinate system) to generate a time series registration coordinate set to ensure that the data from different sensors are effectively connected in the same coordinate system, providing accurate time series data for subsequent path recognition. For example, if the data interval provided by the lidar is 0.1 seconds, the visual sensor is 0.05 seconds, and the inertial navigation unit is 0.2 seconds, the timing can be aligned through interpolation to ensure that the same vehicle state can be corresponded at the same time point; Through timestamp synchronization, linear interpolation, spatial pose correction and coordinate normalization, data complementarity of lidar, visual sensors and inertial navigation units is achieved, solving the limitations of a single sensor and providing more reliable environmental perception capabilities for autonomous driving.

[0021] S112: Based on the overlapping area of ​​the obstacle point position and the image edge contour in the time series registration coordinate set, combined with the attitude angular velocity component, the center offset trajectory of the stable area and the position transformation angle of the adjacent frames are extracted to identify the return path and generate the return channel posture parameter group; After the temporal registration coordinate set is generated, path identification is performed based on the overlap between the obstacle point locations in the temporal data and the image edge contours. By spatially mapping the obstacle point cloud data and the image contours in the temporal registration coordinate system, the intersection of the obstacle and the image contour is identified. Image processing algorithms (such as edge detection or contour extraction) are then used to identify the overlap, which represents key portions of the vehicle's return path. This overlap can eliminate single-sensor false detections (such as LiDAR misidentification of light spots or visual misidentification of shadows), ensuring reliable obstacle location. Combined with the attitude and angular velocity data provided by the inertial navigation unit, the offset trajectory of the center of the stable region is further analyzed to calculate the vehicle's stable offset and the transformation angle between adjacent positions in each frame. For example, when the edge contour of an image changes within a certain time interval, the return path can be obtained by calculating the trajectory change of a specific point in the image in space, and information can be extracted to identify the return path. For example, assuming that the position of the obstacle point changes little within a certain time period and the transformation angle of adjacent frames is small, it means that the vehicle is traveling along a fixed path, thereby identifying the return path. The return path is essentially the inverse process of the outbound path. The motion trajectory of the vehicle when it comes is restored by stabilizing the center offset trajectory of the area. Based on the information, the posture parameter group of the return channel is generated, covering detailed data such as path position and angle, to ensure the accuracy and consistency of the return path.

[0022] The contents of the return channel posture parameter group are shown in Table 1, for example.

[0023] Table 1: Return channel posture parameter group information table

[0024] S113: Based on the pose information of each path segment in the return channel pose parameter group, analyze the sensor configuration mapping relationship in the original coordinate system, convert the path points to a unified reference system, perform distance and angle verification on the channel based on path continuity and pose changes, integrate the dynamic trajectory segments and feasible pose range, and generate a return state identification matrix; Based on the path information in the return channel posture parameter group, the sensor configuration mapping relationship in the original coordinate system is analyzed for each path segment. Each path segment consists of multiple path points, and the posture of each path point includes position coordinates and attitude angle data. During the analysis, the original path point information needs to be converted to the unified reference system based on the transformation relationship between the sensor coordinate system and the unified reference system. For example, if the original path point is located in the local coordinate system, the posture of the path point can be converted to the coordinates of the global coordinate system through a known transformation matrix to ensure the uniformity of the path. The path is verified based on the continuity of the path and the posture changes between each path point, including calculating the distance and angle changes of the path segment. For example, when the angle changes between path points are large, it means that the vehicle is performing a turning operation. Special attention should be paid to the impact of this path change on the vehicle's return to home to ensure the smoothness and consistency of the path. The dynamic trajectories of different path segments are integrated, and the impact of path continuity and attitude changes on the return process is analyzed. The information is used to calculate the feasible attitude range, that is, the maximum deviation angle and minimum distance allowed by the path, to further ensure the feasibility of the path. Finally, a return status identification matrix is ​​generated, which includes the distance verification results, angle verification results and attitude change data of each path segment, forming a comprehensive evaluation model to determine whether the return path identification can be successfully completed.

[0025] The generated return state identification matrix includes the target area location, posture information, obstacle characteristics, and environmental boundary information. Among them, the target area location refers to the location coordinates of the target area during the unmanned vehicle's return process. This information is acquired through sensors and accurately mapped to a unified coordinate system to ensure accurate positioning of the target area during the return process. Attitude information: refers to the motion state of the unmanned vehicle, including angle, speed and other data related to the vehicle's posture. This information is mainly provided by the inertial navigation unit. By monitoring the attitude angle and speed, it helps to assess the direction and stability of the unmanned vehicle during the return process. Obstacle characteristics: This refers to information about surrounding obstacles identified by lidar and vision sensors, including the obstacle's spatial location, shape, and relative position to the vehicle's path. This information is crucial for determining whether the vehicle can successfully avoid the obstacle and return to the target area along a safe path. Environmental boundary information: refers to the boundary data around the vehicle, including environmental restrictions such as the edge position of the road. The data helps determine the passability of the path and ensures that the unmanned vehicle returns along the safest and most appropriate path.

[0026] See also Figure 3 The specific steps for obtaining the traversable return path interval are as follows: S211: Based on the return state identification matrix, the output data of the positioning sensor, lidar, and visual sensor are called to extract the positioning trajectory, obstacle distribution, and environmental boundary trend. The obstacle position is identified based on the obstacle distribution. The path segments affected by the obstacles in the positioning trajectory are detected, and the path segments overlapping with the obstacles are eliminated to generate an obstacle-free passage trajectory. Based on the return state recognition matrix, the positioning trajectory data contained in the matrix is ​​first called and operated. The data is a fusion of lidar, visual sensor and inertial navigation system, and its data format is a coordinate sequence and a corresponding timestamp. The program first reads the trajectory coordinate sequence within the corresponding period. Each trajectory point contains three-dimensional spatial coordinates (x, y, z) and attitude information (θ, φ, ψ). The obstacle area is extracted by combining the point cloud information output by the lidar. The extraction action constructs a scanning circle with a radius of R (set to 3 meters) around the current position of the vehicle and analyzes the point cloud density within the circle. If the point cloud density of a certain area exceeds the threshold ρ (set to more than 150 points per square meter), it is determined to be an obstacle area. For example, the current position of the vehicle is (10, 20), and the obstacles in front are concentrated at (12, 21). The point cloud density is 180 points / ㎡, which meets the obstacle judgment condition. Then, combined with the image processing results of the visual sensor, the Canny algorithm is called to extract the image edge, and then the outer contour line segment of the obstacle in the image is calibrated by the connected area, which is compared with the obstacle extracted by the lidar. The coordinates of the obstacle information are matched. If the distance difference between the two is less than 0.3 meters, they are considered to be the same obstacle. Through this process, multiple sets of obstacle position sets are identified. Then, the path segment impact detection operation is performed on the positioning trajectory. The entire trajectory is segmented at 1-meter intervals. A detection circle with a radius of 0.5 meters is constructed for the center point of each trajectory segment. The distance between the detection circle and the obstacle center point is determined. If it is less than 0.5 meters, the trajectory segment is considered to be affected by the obstacle. For example, the trajectory segment center is (13, 22) and the obstacle position is (13.2, 22 .1), the distance is 0.223 meters, which meets the overlapping condition; after identifying all path segments affected by obstacles, the path removal action is performed on them, that is, the segment is marked as impassable and deleted from the trajectory. This step is implemented through the Boolean mask method. The mask is "0" for removal and "1" for retention. Finally, all path segments not covered by obstacles are retained as barrier-free passable trajectory data. For example, if the initial trajectory has 20 segments and the affected segments are 3, 5, and 7, then the barrier-free trajectory consists of 17 segments excluding these three segments.

[0027] S212: Calling the barrier-free passage trajectory, combining it with the boundary detection data of the visual sensor, identifying the forward angle segment of each trajectory segment, extracting the angle information, calculating the passage space optimization value of the segment, and screening out the optimal passage angle segment; Using the formula: ; in, represents the optimized value of the passage space of the section, represents the trajectory length of the i-th segment, represents the forward angle of the i-th segment, represents the obstacle distribution distance within the i-th section, represents the minimum safety distance of the i-th segment boundary, Represents the total number of sections; The optimized value of the passage space of a section is an important indicator for evaluating the safety and feasibility of the passage of an unmanned vehicle in a specific trajectory section. This value comprehensively considers four key factors: trajectory length, forward angle, obstacle distribution distance, and boundary safety distance. After normalizing the factors, the calculated optimized value can intuitively reflect the passage capacity of the trajectory segment in the current environment. The larger the optimized value, the wider the passage space and the higher the safety of the section. In the calculation process, the trajectory length reflects the ductility of the path, the forward angle is used to measure the rationality of the path orientation, the obstacle distribution distance is used to assess potential risks, and the boundary safety distance indicates the stability of the trajectory segment at the edge of the road. After combining the four, the optimized value obtained by weighted calculation can provide a scientific basis for the selection of the automatic return path of the unmanned vehicle with the support of multi-sensor fusion data. When the optimized value is large, it indicates that the passage conditions in the section are relatively ideal, and the unmanned vehicle can give priority to this section as the forward direction. When the optimized value is small, it indicates that the passage is restricted, there are problems such as dense obstacles or narrow space, and the path should be avoided or replanned. Combined with the boundary detection data of the visual sensor, the forward angle segment of each trajectory segment is identified, and the trajectory length of each segment is determined based on the sensor fusion data. , extract the forward angle of each segment , call the laser radar and visual sensor to measure the length of the trajectory segment in real time, and calculate the length of each segment through path integral. For example, if the sensor data shows that the straight-line distance from the start point to the end point of a segment is 12 meters, then , Secondly, call the visual sensor and lidar to measure the distance of obstacle distribution , calculate the difference between the obstacle coordinates and the current position of the vehicle to obtain the closest distance of the obstacle in each section. For example, in this section, the obstacle distance is 4 meters, then , call the visual sensor to determine the minimum safe distance of the boundary , determined by the distance between the boundary detection line and the vehicle outline. For example, if the boundary safety distance is 1.5 meters, then , the numerical values ​​need to be dimensionally unified and normalized before calculation to ensure the accuracy of the calculation results; Track length Normalization: Maximum trajectory length As a benchmark, divide the length of each segment by For example, assuming ,but ; Obstacle distance Normalization: Maximum detectable distance As a baseline, assuming ,but ; safe distance Normalization: Maximum boundary safety distance As a baseline, assuming ,but ; Forward angle Use radians and no normalization is required, for example Convert to radians ; After completing the above normalization, substitute into the formula: ; This value indicates that the optimized value of the passage space of the current trajectory segment is 0.6196. The larger the value, the more suitable the segment is for passage. After calculating the optimized value of the passage space for all segments, the segment with the largest value is selected as the optimal passage angle segment.

[0028] S213: Calling the optimal pass angle section, combining the target homing point, calculating the angle deviation between the target direction and the section direction, selecting the optimal path segment based on the deviation, and generating a passable return path interval; Determine the coordinates of the target homing point, call the current position of the vehicle, and determine the current direction of the vehicle through multi-sensor data fusion, including data from lidar and visual sensors, for real-time monitoring of the vehicle's surrounding environment to ensure the safety of path selection. Calculate the direction of the line between the vehicle's current position and the target homing point as the target direction, and compare the direction information of each section in the optimal passage angle section with the target direction. Calculate the angle difference between the direction of each section and the target direction one by one. All angle differences are expressed in absolute value to ensure that the deviation value does not appear in positive or negative directions. Arrange all angle deviations in ascending order, and select the direction with the smallest deviation as the preferred direction. In this example, if the vehicle's current position is (3, 4) and the target homing point is (10, 8), then the vehicle's target direction is from (3, 4) to (10, 8). The angle of the connecting line is calculated to be 30 degrees according to the Pythagorean theorem. Multiple directions included in the optimal pass angle segment, such as 25 degrees, 32 degrees, and 40 degrees, are called, and the deviations from the target direction are calculated to be 5 degrees, 2 degrees, and 10 degrees, respectively. Finally, the 32-degree direction with the smallest deviation is selected as the preferred direction. Based on the preferred direction, the route is extended from the vehicle's current position along this direction to the target homing point, and each adjacent pass angle segment is connected in turn to finally generate a passable return path interval, ensuring that the vehicle can safely return from its current position to the target homing point.

[0029] See also Figure 4 , the steps for obtaining the return tolerance control limit are as follows: S311: Based on the traversable return path interval, extract the lateral distance between the vehicle trajectory and the interval centerline, identify the lateral distance value and time interval between the current frame and the previous frame, analyze the lateral distance change trend, determine the lateral change direction and value intensity, and obtain the lateral offset trend value; Obtain the lateral distance information between the trajectory line and the center line of the interval, extract the lateral position data of the vehicle in the current frame and the previous frame. The relative coordinate value of the position data can be obtained by integrating the GNSS and inertial navigation systems. Taking the center of the vehicle body as the reference benchmark, the lateral distance between the trajectory and the center line of each frame is calculated by coordinate difference. The trend of the change of the lateral distance needs to be obtained continuously based on multiple frames of data. The time window is set to 0.2 seconds, and 5 frames of data are sampled. The trend fitting processing of the lateral distance difference sequence is performed. The current direction of movement and trajectory deviation speed of the vehicle are judged by the positive and negative difference and the rate of change of the value. The current trend change value is compared with the lateral change average extracted in the previous stable driving stage. The average rate is compared to determine whether there is a sudden change in the trajectory deviation trend. If the trend change rate is more than twice the previous average rate, the vehicle is considered to be in an offset acceleration state. If the trend change rate is less than half of it, it is judged to be in a stabilizing state. Taking a vehicle driving on a certain road section as an example, the lateral distance changes from 0.15 meters to 0.45 meters in the first 3 seconds, with a change rate of 0.1 meters per second, and the change rate in the current 0.2-second time window reaches 0.25 meters per second, then the acceleration deviation condition is met. Based on the above trend judgment result, it is further assigned to the lateral deviation trend value, expressed in meters per second, for subsequent feasibility quantification processing.

[0030] S312: The lateral offset trend value is combined with the attitude angle change, the obstacle boundary distance extracted from the LiDAR lateral space characteristics, and the trajectory boundary distance to perform a joint operation to calculate the return path feasibility quantification value. Based on the comparison result with the lateral space threshold, the feasibility of the trajectory is determined, and the error band boundary of the passable interval is selected to generate the return tolerance control limit. Using the formula: ; in, Represents the quantitative value of the feasibility of the return path, Represents the horizontal deviation trend value, Represents the trajectory boundary distance value, Represents the change in attitude angle, Represents the obstacle boundary distance value, Represents the trend direction value of the horizontal distance change, Represents the attitude angle change rate; The return path feasibility quantification value is a comprehensive numerical indicator used to quantify whether the unmanned vehicle's path has the predictability of safe passage after combining the attitude angle change and lateral space constraints under the current trajectory deviation state. This value is obtained by integrating key physical quantities such as lateral deviation trend, trajectory and obstacle boundary distance, attitude change angle and change rate, and performing combined operations after unified normalization. Its numerical value reflects the degree of adaptation of the current trajectory within the spatial tolerance range. The closer the value is to zero, the more sufficient the path space and the higher the attitude stability, and the stronger the feasibility of passage. The closer the value is to or exceeds the preset threshold, the greater the risk of intense deviation trend, approaching obstacles or unstable attitude, which means that the passage space is restricted. Therefore, this value can be used as a key judgment basis for dynamically determining whether the path meets the return passage requirements; Combined with the vehicle attitude angle change, the obstacle boundary distance value in the laser radar lateral space feature and the trajectory boundary distance value, a joint calculation is performed. First, each participating item is normalized and the dimension is unified to avoid the imbalance of calculation scale caused by different physical quantities. In the common domestic road design standards, the conventional vehicle width is set to 2 meters, so this value is selected as the normalized reference length. All distance-related parameters are divided by this value to complete the normalization. When obtaining the lateral offset trend value A, the lateral coordinate sequence of the vehicle's trajectory line in a continuous time period is collected. With 0.2 seconds as the time window and a sampling frequency of 10Hz, 5 frames of data are extracted, and the lateral difference calculation between frames is performed and the mean statistics are performed. For example, the lateral coordinates of the vehicle in the 5 frames are 1.25 meters, 1.35 meters, 1.50 meters, 1.75 meters, and 2.00 meters, respectively. Then, its lateral change rate is 0.10, 0.15, 0.25, and 0.25 meters per second, respectively, with an average of 0.1875 meters per second. After normalization, A = 0.09375. The trajectory boundary distance L is composed of the maximum distance from the center of the vehicle to the effective boundary of the road. The left and right boundary point sets of the trajectory at the corresponding time are extracted from the lidar scanning interval. L = 1.8 meters is extracted through nearest neighbor judgment, and L = 0.9 is normalized. The attitude angle change θ is obtained by integrating the IMU angular velocity. For example, if the three-axis angular velocity acquisition values ​​are 0.02, 0.03, and 0.01 radians per second, the integration within a 0.2-second time window yields Q = 0.012 radians. The obstacle boundary distance P is calculated from the point with the maximum distance intensity in the lidar scan. In this example, the distance from the obstacle point to the center of the vehicle is 0.6 meters, and the normalization yields P = 0.3. The lateral direction R is determined by the positive or negative lateral offset trend. The current offset trend value is positive, so R = 1. The attitude angle change rate M is obtained by comparing the difference between the attitude angle changes of the previous and next frames to time. For example, if the attitude angle changes from 0.05 to 0.08 radians in 0.5 seconds, M = 0.06 radians per second. Substituting the normalized data into the following: A = 0.09375, L = 0.9, Q = 0.012, P = 0.3, R = 1, M = 0.06; The first part of the calculation: ; The second part of the calculation: ; The final result is: ; The following are the parameters: A represents the normalized lateral deviation trend value, expressed in units relative to the standard vehicle width; L represents the normalized trajectory boundary distance value, based on a maximum spatial width of 2 meters; Q represents the attitude angle change, expressed in radians, obtained by integrating the angular velocity; P represents the normalized obstacle boundary distance value; R represents the lateral motion direction, with a positive value indicating a right deviation; M represents the attitude angle change rate, expressed in radians per second; and Y represents the normalized return path feasibility quantification value. A larger value indicates an increased risk of spatial trend deviation. By normalizing and integrating the various quality criteria into a unified structure, the accuracy and consistency of the return trajectory assessment were improved. The results showed that the feasibility quantification value of the return path was 0.7470, which was lower than the baseline traversable safety threshold of 1.0. Therefore, this trajectory was determined to be a traversable area and was further used to screen the trajectory error band boundaries and ultimately establish the return tolerance control limit.

[0031] See also Figure 5 , the steps for obtaining the steering input adjustment instruction set are as follows: S411: Invoke the return-to-home tolerance control limit, extract the steering angle input and the yaw angle calculated by the fusion inertial navigation and vehicle attitude sensor, and determine the sign consistency and angle difference. If the direction does not match or the angle difference exceeds the threshold, the return-to-home control deviation is determined and the attitude control deviation status is obtained; The vehicle position and attitude solution obtained by fusion of GNSS and IMU in the multi-sensor fusion platform is used as the reference state information, and the yaw angle of the current frame is extracted as the vehicle attitude direction reference. The original steering angle input value issued by the control configuration is simultaneously obtained as a comparison item. The direction consistency is determined by the direction sign, that is, whether the yaw angle value and the steering angle value have the same sign. If there is a sign inconsistency, the directions are considered to be opposite. At the same time, the absolute angle values ​​of the two are extracted to calculate their angle offset. The offset calculation formula is the absolute value of the difference between the two angle values. The offset is then compared with the tolerance angle base value set in the vehicle system. The quasi-threshold is used for judgment and comparison. The threshold is set according to the vehicle structure design and dynamic characteristics. For example, for a medium-sized four-wheeled unmanned vehicle, the angle deviation tolerance is set to 5 degrees, which is equal to 0.087 radians. When the current offset angle is greater than the threshold, it is determined that there is a significant deviation between the current attitude response direction of the vehicle and the control command direction. This judgment serves as the basic basis for the steering control deviation attitude response. In the application scenario, if the unmanned vehicle performs a return turn in a narrow channel, if the control configuration issues a left turn command but the vehicle's yaw direction still maintains a rightward deviation trend, immediate feedback is required. This process obtains the attitude control deviation state.

[0032] S412: Based on the attitude control deviation state, combined with the road edge offset detected by the visual sensor and the steering feedback angle, the steering input amplitude is dynamically adjusted through multi-sensor data fusion. The return trajectory steering correction command is calculated and fed back to the vehicle control system. The original control variable is superimposed and updated in real time to obtain the steering input adjustment command set. Using the formula: ; in, Indicates the return trajectory steering correction instruction, Indicates the attitude control deviation state quantity, Indicates the lateral offset value of the road edge detected by the fusion vision sensor, Indicates the Multi-sensor feedback correction factor, Indicates the current steering feedback angle, is the number of sensor feedback factors; The return trajectory steering correction instruction refers to a steering angle correction value obtained through comprehensive calculation based on information such as the current vehicle posture deviation state, the visual boundary offset of multi-sensor fusion feedback, the front wheel feedback angle, and multiple dynamic posture response factors during the unmanned vehicle's automatic return process. It is used to make real-time dynamic adjustments to the original steering control input, so that the vehicle's driving trajectory gradually returns to the set return path. The instruction is expressed in radians and can quantify the steering amplitude that needs to be adjusted. The larger the value, the more serious the vehicle's current deviation from the trajectory, and the greater the steering compensation required. A value close to zero means that the posture is basically consistent with the path. As a key component of the steering input adjustment instruction set, this instruction directly participates in the control system's return trajectory correction control. The steering input correction operation is performed by combining the key feedback extracted from the multi-sensor fusion platform. First, the real-time output image data of the visual sensor is obtained and the lane boundary model is analyzed. The lateral distance difference between the geometric center of the vehicle body and the center of the lane is calculated, which is defined as the boundary offset displacement. This value is processed with the vehicle width as the normalized reference. Assuming that the standard width of the vehicle is 1.8 meters, the current detection distance is 0.6 meters on the left and 0.8 meters on the right, the center offset of the vehicle body is , its dimensionless unit is converted to a dimensionless ratio relative to the vehicle width, and then the front wheel feedback angle value is obtained from the vehicle control configuration The servo angle fed back by the electronic power steering module is currently 0.09 radians, and the yaw rate value is obtained by the IMU as , currently 0.04 radians per second, the vehicle side slip tendency response value is obtained by detecting the left and right wheel speed difference through the wheel speed sensor and set as The front wheel lateral deviation is calculated by the front axle torque model as The above three attitude feedback quantities are normalized to 0.2, 0.15, and 0.25 respectively based on the maximum attitude change limit of 0.2 radians per second, and then the attitude control deviation state quantity is taken The deviation angle obtained by comparing the control angle input value with the actual yaw angle direction is 0.1 radians, which is normalized to 0.2 radians. ; in: ; The calculation process is as follows: ; ; The final calculated return trajectory steering correction command is 0.4216 radians, indicating the steering angle that needs to be adjusted after the current fusion multi-source deviation correction. This value will be superimposed on the current control command to correct the return trajectory in real time and establish the steering input adjustment command set; in, Indicates the return trajectory steering correction instruction, Indicates the attitude control deviation state, which is based on the normalized result of the difference between the steering input angle and the yaw direction. represents the lane boundary offset detected by the visual sensor, normalized to the vehicle width ratio, Indicates the The attitude feedback factors include yaw rate, body side slip response, and front wheel side deviation. Indicates the front wheel feedback angle value, Represents three posture feedback factors. The purpose of this formula is to normalize and integrate visual displacement, posture feedback and direction error to form a highly responsive steering correction strategy. The obtained correction angle value is 0.4216 radians, which is within the controller's allowable correction range of 0-0.6 radians and can be used as the core component of the steering input adjustment instruction set.

[0033] See also Figure 6 The specific steps for obtaining the return path maintenance status tag are as follows: S511: Based on the steering input adjustment instruction set, the overlapping area between the vehicle trajectory output by the positioning unit and the path detected by the vision unit is collected, the lateral offset and overlap length of the trajectory are extracted, the variation amplitude of the overlap length within the period is analyzed, and a periodic overlap amplitude sequence is generated; This is achieved through the fusion of multiple sensors, such as LiDAR and camera-based visual positioning. In practice, the positioning unit's trajectory is calibrated using GPS signals and an inertial navigation system. The visual unit processes video images in real time to acquire the road path and extract two key parameters: lateral offset and overlap length. Lateral offset represents the deviation between the vehicle's current position and the ideal path, while overlap length represents the length of overlap between the vehicle's trajectory and the detected path. During this process, lateral offset is calculated in real time through real-time trajectory data updates. At each point in time, the trajectory is compared with the ideal path to determine the current offset. Overlap length is determined by scanning the detected area of ​​the actual road and, combined with image analysis techniques, extracting the length of the overlap area. Analyzing the variation in overlap length within a cycle involves time series analysis. By continuously observing the raw data, the range between the maximum and minimum overlap length values ​​within each cycle is calculated, reflecting the stability of the path. This generates a periodic overlap amplitude sequence. In practical applications, this data can help assess the accuracy and stability of path tracking, preparing for further analysis.

[0034] S512: Calling the periodic overlapping amplitude sequence, extracting the overlapping amplitude difference sequence within consecutive periods, marking the stable trend according to the polarity of the difference change, and combining it with the set visual trajectory stability threshold to determine whether the overlapping change converges to a single trend, and obtain the visual trajectory fusion stable trend value; The overlap amplitude differences within consecutive cycles are extracted from the periodic overlap amplitude sequence. By comparing the overlap amplitudes of two consecutive cycles, the difference between each pair of cycles is calculated. This difference reflects the trend of path overlap. Based on the polarity of the difference changes, a stable trend is identified. Specifically, when the difference changes tend to stabilize, the path overlap is considered stable. This process relies primarily on polarity analysis within a mathematical model. For example, the difference sequence for each cycle is differentiated to obtain the rate of change. When the rate of change approaches zero, a stable trend is identified. Based on this, a set visual trajectory stability threshold is used to further determine whether the overlap changes have converged to a single trend. The visual trajectory stability threshold is determined by a preset maximum allowable deviation value. The threshold can be derived from raw data or actual testing. For example, the threshold is set to 0.05 meters. When the difference change falls below this value, the trajectory is considered stable. This process ensures that trajectory stability can be determined. In practical applications, if the trajectory change amplitude remains below the threshold, the path tracking is considered stable, and the visual trajectory fusion stability trend value is obtained. This is accomplished by determining whether the overlapping changes in the paths gradually converge to a single trend. If the overlapping changes eventually converge to a single trend, the visual trajectory fusion stable trend value is the value of that trend. This means that when the trajectory changes continue to be small and tend to be stable, the path tracking state can be considered stable. This stable trend value becomes an important indicator for determining whether the path is effective and stable.

[0035] S513: Based on the visual trajectory fusion stability trend value, the lateral difference between the inertial navigation pose solution and the visual positioning coordinates within the stability period is identified. Combined with the path overlap area length ratio, it is determined whether the unmanned vehicle maintains stable path tracking and outputs the return path maintenance status label; Identify the lateral difference between the inertial navigation pose solution and the visual positioning coordinates within the stable period. This process first requires obtaining the pose data provided by the inertial navigation system (INS) and comparing it with the positioning coordinates calculated by the vision unit. By comparing the lateral deviation between the two, the difference value is obtained. The difference value reflects the error size of the vehicle during actual driving. Combined with the path overlap area length ratio, the relationship between the lateral difference and the overlap area length ratio is calculated. The path overlap area length ratio is obtained by calculating the ratio of the length of the overlap area to the total length of the entire path. This ratio is used to characterize the integrity and stability of the path. When the ratio is high, it means that the path is more stable. By analyzing the ratio of the lateral difference to the path overlap area length, it is possible to determine whether the unmanned vehicle maintains stable path tracking and output a return path maintenance status label (when determining whether the unmanned vehicle maintains stable path tracking, the return path maintenance status label should be output when the stable path tracking is maintained, rather than outputting the label regardless of whether the stable path tracking is maintained. This means that the return path maintenance status label needs to be updated or output only when the unmanned vehicle successfully maintains stable path tracking. If the unmanned vehicle does not maintain stable path tracking, the label should not be output). The label is used to determine whether the vehicle can continue to perform the return mission. In actual applications, assuming that the lateral difference is 0.03 meters and the path overlap area length ratio is 0.95, it is determined that the unmanned vehicle's path tracking is stable and can continue to perform the return mission; Table 2: Test data example: path overlap, difference sequence and stability analysis

[0036] As shown in Table 2 , the variation in the overlap amplitude within a cycle is always within a small range (between 0.8 and 0.85 m), while the rate of change in the difference is always kept at a low interval (e.g., 0.05 m / cycle), proving that the path remains stable, the lateral difference remains within 0.03 m, and the ratio of the path overlap area is always 0.95, indicating that the overlap area of ​​the path is relatively stable and has strong stability.

[0037] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. The automatic return method of unmanned vehicle multi-sensor data fusion is characterized by: The following steps are involved: S1: Obstacle point information is acquired through lidar, image edge contours are acquired through visual sensors, and attitude data is acquired through inertial navigation units. The three types of data are synchronized and registered in time and space, and the state of the target area associated with the return path is extracted. The spatial mapping model between sensor devices is identified and the reference coordinate system is unified to generate the return state identification matrix. S2: Based on the return state identification matrix, the positioning trajectory, obstacle distribution and image boundary trend are extracted, the forward angle segment of the passable space is screened, and the optimal travel direction is selected in combination with the target homing point to generate a passable return path interval; S3: Based on the traversable return path interval, by analyzing the lateral distance trend between the vehicle trajectory line and the interval centerline, combining the attitude angle change to determine the trajectory offset and amplitude, and combining the lateral spatial characteristics of the LiDAR to determine the feasibility of passage, screen the effective error band boundary, and generate the return tolerance control limit; S4: calling the return tolerance control limit, comparing the steering control input with the current attitude deflection direction, and adjusting the steering angle direction and amplitude if there is a deviation, and correcting the control instruction according to the visual sensor feedback to obtain the steering input adjustment instruction set.

2. The automatic return method for unmanned vehicle multi-sensor data fusion according to claim 1 is characterized in that: The return state identification matrix includes the target area location, attitude information, obstacle characteristics, and environmental boundary information; the navigable return path interval includes the navigable area, obstacle distribution, travel direction, and boundary contour; the return tolerance control limit includes the lateral distance threshold, attitude angle threshold, lateral space tolerance, and error band width; and the steering input adjustment instruction set includes the steering angle, steering direction, and steering correction instruction.

3. The automatic return method for unmanned vehicle multi-sensor data fusion according to claim 1 is characterized in that: The steps for obtaining the return status identification matrix are specifically as follows: S111: Obstacle point information is acquired through the lidar, image edge contours are acquired through the visual sensor, and attitude data is acquired through the inertial navigation unit. The data are jointly interpolated using timestamps to extract the spatial pose offset and sensor imaging time series alignment error of each time slice. The position information of the three types of data is used for coordinate normalization processing to generate a time series registration coordinate set. S112: Based on the overlapping area of ​​the obstacle point position and the image edge contour in the time-series registration coordinate set, combined with the attitude angular velocity component, extract the stable area center offset trajectory and the adjacent frame position transformation angle, identify the return path, and generate the return channel posture parameter group; S113: According to the posture information of each path segment in the return channel posture parameter group, the sensor configuration mapping relationship in the original coordinate system is analyzed, the path points are converted to a unified reference system, the channel is verified for distance and angle based on the path continuity and posture change, the dynamic trajectory segment and the feasible posture range are integrated, and a return state identification matrix is ​​generated.

4. The automatic return method for unmanned vehicle multi-sensor data fusion according to claim 3 is characterized in that: The steps for obtaining the passable return path interval are specifically as follows: S211: Based on the return state identification matrix, the output data of the positioning sensor, the lidar, and the visual sensor are called to extract the positioning trajectory, obstacle distribution, and environmental boundary trend. The obstacle positions are identified according to the obstacle distribution. The path segments affected by the obstacles in the positioning trajectory are detected, and the path segments overlapping with the obstacles are eliminated to generate an obstacle-free passage trajectory. S212: Calling the barrier-free passage trajectory, combining it with the boundary detection data of the visual sensor, identifying the forward angle segment of each trajectory segment, extracting the angle information, calculating the passage space optimization value of the segment, and screening out the optimal passage angle segment; S213: Call the optimal pass angle section, combine it with the target homing point, calculate the angle deviation between the target direction and the section direction, select the optimal path segment according to the deviation, and generate a passable return path interval.

5. The automatic return method for unmanned vehicle multi-sensor data fusion according to claim 4 is characterized in that: The steps for obtaining the return tolerance control limit are as follows: S311: Extracting the lateral distance between the vehicle trajectory and the centerline of the navigable return path interval, identifying the lateral distance value and time interval between the current frame and the previous frame, analyzing the lateral distance change trend, determining the lateral change direction and numerical intensity, and obtaining a lateral offset trend value; S312: Call the lateral offset trend value, combine it with the attitude angle change, the obstacle boundary distance and the trajectory boundary distance extracted from the laser radar lateral space feature, perform a joint operation, calculate the feasibility quantification value of the return path, judge the feasibility of the trajectory based on the comparison result with the lateral space threshold, and screen the error band boundary of the passable interval to generate the return tolerance control limit.

6. The automatic return method for unmanned vehicle multi-sensor data fusion according to claim 5 is characterized in that: The steps for obtaining the steering input adjustment instruction set are specifically as follows: S411: Invoking the return-to-home tolerance control limit, extracting the steering angle input and the yaw angle calculated by the fusion of the inertial navigation and the vehicle attitude sensor, performing sign consistency and angle difference determination, and if the direction does not match or the angle difference exceeds a threshold, determining the return-to-home control deviation and obtaining the attitude control deviation status; S412: Based on the attitude control deviation state, combined with the road edge offset and steering feedback angle detected by the visual sensor, the steering input amplitude is dynamically adjusted through multi-sensor data fusion, and the return trajectory steering correction instruction is calculated. The instruction is fed back to the vehicle control, superimposed on the original control amount, and updated in real time to obtain a steering input adjustment instruction set.

7. The automatic return method for unmanned vehicle multi-sensor data fusion according to claim 1 is characterized in that: The method further comprises step S5: S5: Based on the steering input adjustment instruction set, collect the vehicle trajectory output of continuous cycles and the change in the length of the overlapping area of ​​the path line in the visual image, analyze whether the change trend is in a convergence state, determine the stability period after the control is executed, and output the return path maintenance status label; The return path maintenance status label includes track overlap rate, path stability indicator, and visual path consistency.

8. The automatic return method for unmanned vehicle multi-sensor data fusion according to claim 7 is characterized in that: The steps for obtaining the return path holding status tag are specifically as follows: S511: Based on the steering input adjustment instruction set, collecting the overlapping area between the vehicle trajectory output by the positioning unit and the path detected by the vision unit, extracting the lateral offset and overlapping length of the trajectory, analyzing the variation amplitude of the overlapping length within the period, and generating a periodic overlapping amplitude sequence; S512: Calling the periodic overlapping amplitude sequence, extracting the overlapping amplitude difference sequence within consecutive periods, marking the stable trend according to the polarity of the difference change, and combining the set visual trajectory stability threshold to determine whether the overlapping change converges to a single trend, thereby obtaining the visual trajectory fusion stable trend value; S513: Based on the visual trajectory fusion stability trend value, the lateral difference between the inertial navigation pose solution and the visual positioning coordinates within the stable period is identified, and combined with the ratio of the length of the path overlap area, it is determined whether the unmanned vehicle maintains stable path tracking, and the return path maintenance status label is output.

Images