Unmanned system navigation line sensing and obstacle distance measuring method in weak texture scene

Abstract

The invention discloses an unmanned system navigation line sensing and obstacle distance measuring method in a weak texture scene, and belongs to the technical field of machine vision. The method comprises the following steps: firstly, detecting a reference object and a pedestrian in an image by using an improved YOLOv5 network to obtain reference object bounding box coordinate information and pedestrian detection box coordinate information; extracting an initial reference point set, removing outliers through an RANSAC (Random Sample Consensus) algorithm, fitting a row line by using a least square fitting algorithm, and extracting a navigation line; meanwhile, region division is carried out on the target scene, the minimum depth value of each region is calculated, the approximate minimum depth of the pedestrian is determined through a four-partition depth comparison algorithm, then the depth range of the pedestrian is obtained, and distance information of the pedestrian is calculated; and finally, through fusion of improved YOLOv5 target detection and binocular vision technologies, extraction of navigation lines and distance measurement of pedestrian obstacles are realized, so that environment perception precision of unmanned system navigation, obstacle distance measurement accuracy and system real-time response capability are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine vision, and specifically relates to a method for unmanned system navigation line perception and obstacle ranging in a weak texture scene. Background Art

[0002] With the development of intelligent engineering, the demand for autonomous operation of unmanned systems is becoming increasingly urgent, and its core technology depends on the accuracy and robustness of the environmental perception and positioning system. Currently, unmanned systems face the following technical bottlenecks in complex unstructured environments: Traditional methods rely on artificially preset road edges or visual detection technologies based on fixed threshold segmentation. However, in the row-by-row scenario, due to the lack of clear road boundaries and problems such as foliage occlusion and uneven illumination, the cumulative deviation of navigation line fitting occurs, making it difficult to meet the real-time path tracking requirements of unmanned systems; Existing ranging algorithms based on binocular stereo matching rely on high-precision depth calculation. However, in scenarios where pedestrian postures are variable and distances change dynamically, traditional methods are difficult to quickly segment the pedestrian area due to the lack of combination of target semantic information, resulting in increased ranging delay and error, affecting the obstacle avoidance response speed. To address the above problems, existing technologies have tried to improve the perception ability by improving the target detection algorithm, but there are the following limitations: The detection method based on YOLOv5 does not optimize the selection strategy of the reference point of the detection frame, and is easily affected by occlusion when directly using the center point of the detection frame to fit the row line; The pedestrian ranging algorithm does not combine the depth distribution characteristics of the detection frame, and relying only on a single point depth value easily leads to fluctuations in the ranging results. Summary of the Invention

[0003] To overcome the deficiencies in the traditional unmanned system navigation, such as the difficulty in extracting navigation lines between rows due to the obvious dependence of environmental perception on road edge lines, and the low accuracy and poor real-time performance in obstacle ranging, and aiming at the problem of reduced accuracy caused by the inability of the center point of the pedestrian detection bounding box to correspond to the center point of the pedestrian target due to the fuzzy row-by-row features and diverse pedestrian postures, the present invention provides a method for unmanned system navigation line perception and obstacle ranging in a weak texture scene. High-precision detection is achieved according to the trained optimal YOLOv5 convolutional neural network model, the row line is fitted by combining RANSAC and the least squares method to generate a navigation path, and a four-zone depth comparison algorithm is innovatively adopted to approximately segment the image and calculate the depth of pedestrians. This method significantly improves the environmental perception accuracy, obstacle ranging accuracy, and system real-time response ability of unmanned system navigation.

[0004] Technical Solution: To achieve the above objectives, the present invention adopts the following technical solutions:

[0005] The method for unmanned system navigation line perception and obstacle ranging in a weak texture scene of the present invention includes the following steps: S1: Use the YOLOv5 detection model to perform real-time detection on the input image, obtain the coordinate information of all reference bounding boxes in the image, and extract the initial reference point set based on the midpoint at the bottom of the reference detection box; S2: Perform outlier detection and removal on the initial reference point set obtained in step S1: Based on the RANSAC algorithm, for different detection situations, distinguish inliers and outliers, identify and remove the outliers that deviate from the normal distribution in the point set, and obtain the reference point set; S3: Use the least squares fitting algorithm to perform least squares fitting on the reference point set obtained in step S2, obtain the left and right lane lines respectively, and calculate and extract the final navigation line again through the geometric relationship of the fitted left and right lane lines based on the least squares fitting; S4: Obtain the target scene with reference to the final navigation line in step S3, divide the target scene into regions, calculate the minimum depth value of each region, compare the minimum depth values of each region, and determine the approximate minimum depth of the pedestrian; S5: Use the minimum depth obtained in S4 to construct a depth range model, obtain the depth range of the pedestrian, segment the depth image, calculate the average depth within the depth range of the pedestrian, and obtain the distance information of the pedestrian according to the average depth.

[0006] Further, the obtaining of the coordinate information of all reference bounding boxes in the image described in step S1 is achieved by geometric analysis of the target detection framework to extract the coordinate information of the reference bounding boxes ( , , , ), where, ( , ) is the center point coordinate of the reference detection box, , are the width and height of the reference detection box respectively. The specific method is as follows: S1-1: Based on the geometric characteristics of the root point of the input image, define the midpoint at the bottom of the detection bounding box as the reference point for lane line fitting ( , ), and the calculation formula is: , S1-2: Input the image to be detected into the YOLOv5 detection model, output the set of image detection bounding boxes through feature extraction, extract the midpoint coordinates at the bottom of each detection box, obtain the initial reference points, and obtain the initial reference point set.

[0007] Further, the specific method of step S2 is as follows: S2-1: According to the initial reference point set obtained in step S1, and according to sort the reference point set in ascending order of pixel coordinates; calculate the number of the reference point set; if <4. Outlier detection is performed according to the threshold method. Traverse the reference point set to calculate the , difference between the current point and the reference point. If and , then this point is regarded as an outlier; if >4, the RANSAC algorithm is used to remove the outliers to obtain the finally optimized reference point set; where represents the pixel coordinate difference between the current reference point and the previous reference point on the x-axis, represents the preset threshold for judging whether it is abnormal, represents the pixel coordinate difference between the current reference point and the previous reference point on the y-axis, represents the preset threshold for judging whether it is abnormal; S2-2: The calculation steps for RANSAC to obtain inliers are as follows: Set parameters for the initial data point set: predefined maximum number of iterations and error tolerance threshold , and initialize the optimal model parameters; Randomly select initial inliers: Randomly select some data points from the initial data point set as initial inliers; Fitting calculation: Generate a hypothesis model according to the selected initial inliers, calculate the fitting error between all data points in the initial data point set and the current hypothesis model, and record the points with errors less than the threshold t as the current inliers; Update the best model: If the current number of inliers exceeds the historical best record and the model error is lower, update the optimal model parameters, the best error threshold, and the inlier set; Repeat iteration: Repeat the above steps and terminate the calculation when the preset number of iterations is reached; Final inlier set: The finally obtained inlier set is the optimized reference point set.

[0008] Furthermore, the specific method of step S3 is as follows: S3-1: For the left and right line models on the two-dimensional plane, they are represented by the following formula: , where is the slope of the line, is the intercept of the line, , represent the horizontal and vertical coordinates of the fitted line in the reference system. For the given data point , the least squares problem is described as minimizing the sum of squared residuals RSS: , Among them, is the number of data points. By taking the derivative of RSS and setting the derivative to 0, the optimal row line fitting parameters are calculated and estimation values of: , According to the above formula, the slope of the fitted row line and are obtained. Finally, the fitted row line model is ; S3-2: Based on the reference point set processed by RANSAC, perform least squares fitting. After obtaining the left and right row lines respectively, take 10 points symmetrically and equidistantly on the obtained left and right row lines, and denote them as , respectively. Among them , calculate the middle point according to the selected points. The calculation formula is as follows: , Finally, perform least squares fitting on the middle point set to extract the navigation line.

[0009] Furthermore, step S4 is a pedestrian ranging algorithm based on four-zone depth comparison. Using the binocular vision ranging principle, two cameras observe a scene simultaneously. By comparing the differences between the images captured by the two cameras, the depth information of the object is calculated, and then the depth of the object is calculated by triangulation to achieve ranging. When detecting, the detection bounding box is divided into 2*2, that is, 4 zones, and the minimum depth is calculated for each zone and compared. The specific method is as follows: S4-1: First, detect the image of the camera through the YOLOv5 convolutional neural network, and obtain the coordinate information of the pedestrian detection box as . According to the size of the image ( ), restore the pixel coordinates, ( ), which respectively represent the center pixel coordinates of the detection bounding box ( ), and the width and height of the inspection bounding box; S4-2: Then obtain the depth map based on binocular vision from the left and right images. Divide the detection box into four upper, lower, left, and right zones, and perform depth statistics on each zone. For each zone, the depth means from the upper center point to the center point of the zone, from the bottom center point to the midpoint of the zone, from the left center point to the center point of the zone, and from the right center point to the center point of the zone are respectively representing the four zones; S4-3: Extract the depth means on both sides of the center point Take the smaller value between the two for comparison , sort all the depth values on the smaller side, that is Operation, the depth values at the top and bottom are processed in the same way to obtain And sort; S4-4: Finally, for And Compare, select the 5 smallest depth values on the side with the smaller value and take the average as the approximate minimum depth of this area ; Finally, sort the approximate minimum depths of the four areas to obtain , select the smallest depth as the depth of the pedestrian .

[0010] Furthermore, the specific method of step S5 is to define the depth range as -0.2, +0.2] according to the pedestrian depth obtained in S4, and obtain the distance of the pedestrian , and the calculation formula is as follows: , , , .

[0011] Advantageous effects: Compared with the prior art, the present invention realizes the high-precision extraction of navigation lines and the accurate ranging of pedestrian obstacles through the improved YOLOv5 convolutional neural network and binocular vision technology. Experimental results show that the average angular deviation of the navigation line extraction algorithm of the present invention does not exceed 5° in different environments, and the average detection time is about 50 ms, meeting the requirements of navigation line extraction between rows. At the same time, the relative error of the pedestrian ranging algorithm is less than 1% within 6 m, and the average ranging time meets the real-time requirement, providing reliable technical support for the autonomous operation of unmanned systems between rows. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is the overall block diagram of the present invention;

[0013] Figure 2 In (a) is the original image of the orchard scene, (b) is the original image of the wasteland scene, (c) is the navigation line extraction map of the orchard scene, and (d) is the navigation line extraction map of the wasteland scene;

[0014] Figure 3 In (a) is the statistical result of the angular deviation in the orchard scene, and (b) is the statistical result of the angular deviation in the wasteland scene;

[0015] ​Figure 4 Pedestrian pose ranging results at a distance of 1 m for experimental shooting: Figure 4 The measured distance in (a) is 1.0 m, the measured distance in (b) is 1.2 m, and the measured distance in (c) is 1.1 m;

[0016] Figure 5 Pedestrian pose ranging results at a distance of 3 m for experimental shooting: Figure 5 The measured distance in (a) is 2.7 m, the measured distance in (b) is 2.9 m, and the measured distance in (c) is 2.8 m;

[0017] Figure 6 Pedestrian pose ranging results at a distance of 8 m for experimental shooting: Figure 6 The measured distance in (a) is 8.1 m, the measured distance in (b) is 8.0 m, and the measured distance in (c) is 8.2 m. Detailed implementation manners

[0018] The present invention will be further clarified below in conjunction with the accompanying drawings and specific embodiments.

[0019] The present invention provides a method for unmanned system navigation line perception and obstacle ranging in a weak texture scene as Figure 1 shown, which specifically includes:

[0020] S1: Use the YOLOv5 detection model to perform real-time detection on the input image, obtain the coordinate information of all image bounding boxes in the image, and extract the initial reference point set based on the midpoint at the bottom of the image detection box;

[0021] S2: Perform outlier detection and removal on the initial reference point set. Based on the RANSAC (Random Sample Consensus) algorithm, distinguish inliers and outliers for different detection situations, identify and remove the outliers deviating from the normal distribution in the point set, and obtain the reference point set;

[0022] S3: After removing the outliers, use the least squares fitting algorithm to fit the distribution characteristics of the left and right lane lines. Based on the geometric relationship of the left and right lane lines obtained by fitting, extract the final navigation line through least squares fitting calculation again;

[0023] S4: Obtain the target scene with reference to the final navigation line in step S3, divide the target scene into regions, and calculate the minimum depth value of each region. Compare the minimum depth values of each region to determine the approximate minimum depth of the pedestrian;

[0024] S5: Use the minimum depth obtained in S4 to construct a depth range model, obtain the depth range of the pedestrian, segment the depth image, calculate the depth mean value within the depth range of the pedestrian, and obtain the distance information of the pedestrian according to the depth mean value.

[0025] Preferably, the extraction of the reference bounding box coordinate information in step S1 specifically refers to ( , , , ). Among them, ( , ) is the center point coordinate of the image detection box, , is the height of the detection box.

[0026] Based on the geometric features of the image root point, the midpoint at the bottom of the detection bounding box is defined as the initial reference point for line fitting ( , ), and the calculation formula is: ,

[0027] Input the image to be measured into the trained object detection model, output the set of detection bounding boxes through feature extraction, then obtain the initial reference points according to the above formula, and perform outlier rejection through the RANSAC algorithm to obtain the optimized reference point set. For different scenarios, the obtained initial reference points basically match the actual image root points. However, due to the influence of complex factors such as environment and perspective, not all of the obtained initial reference points can be used as reference points for line fitting. The specific problems can be described as:

[0028] Within the image field of view, too few image detection boxes can be detected unilaterally, resulting in a large deviation during line fitting. Therefore, it is necessary to retain as many initial reference points as possible.

[0029] The reference points obtained in adjacent images are actually interference points, which will affect the subsequent line fitting. Therefore, it is necessary to remove the interference points.

[0030] Preferably, for the problems occurring in step S1 in step S2, based on the RANSAC (Random Sample Consensus) algorithm, for different detection situations, distinguish inliers and outliers, remove outliers, and obtain the reference point set.

[0031] S2-1: According to the initial reference point set obtained in step S1, and arrange the reference point set in ascending order according to pixel coordinates; calculate the number of the reference point set; if <4, perform outlier detection according to the threshold method, traverse the reference point set to calculate the , difference between the current point and the reference point. If , and , then consider this point as an outlier; if If it is > 4, the RANSAC algorithm is used to remove the outliers to obtain the finally optimized reference point set; where, represents the pixel coordinate difference of the current reference point and the previous reference point on the x-axis, represents the preset judgment threshold for abnormality, represents the pixel coordinate difference of the current reference point and the previous reference point on the y-axis, represents the preset judgment threshold for abnormality.

[0032] S2-2: RANSAC is a robust model fitting method that filters the effective inlier set from the noisy data through iterative operations and calculates the optimal model parameters.

[0033] The calculation steps for RANSAC to obtain inliers are as follows:

[0034] Set parameters for the initialized data set: predefined maximum number of iterations and error tolerance threshold , and initialize the optimal model parameters. The value of the point is calculated according to this formula, is the desired success probability, is the proportion of inliers in the data set (the smaller the value, the higher the noise in the data, and take 0.4 - 0.6 in the experiment), is the minimum number of samples required for the fitting model (for example, at least two points are required to fit a straight line). The value of is related to the characteristics of the data noise (usually take 1 - 3 in image feature matching) and needs to be adjusted according to the actual distribution.

[0035] ,

[0036] Randomly select inliers: Randomly select some points from the data points as the initial inliers (inliers are the data points after removing the outliers).

[0037] Perform fitting calculation: Fit the model according to the selected points, calculate the fitting error of all data points to the model, and mark the points with an error less than the threshold as the current inliers.

[0038] Update the best model: If the number of current inliers meets the requirements and the fitting error is smaller than the current best error, update the best model and the best error threshold.

[0039] Repeat iteration: Repeat the above steps until the maximum number of iterations is reached and then end.

[0040] Final inlier set: The finally obtained inlier set is the optimized reference point set.

[0041] Preferably, the basic idea of step S3 is to use the least squares fitting algorithm to fit and extract the final navigation line from the distribution characteristics of the lane lines. For the lane line model on a two-dimensional plane, it is represented by the following formula: , where, is the slope of the straight line, is the intercept of the straight line, , represent the abscissa and ordinate of the fitted lane line in the reference system. For a given reference point , the least squares problem is described as minimizing the residual sum of squares (RSS): , where, is the number of data points. By taking the derivative of RSS and setting the derivative to 0, the estimated values of the optimal straight line fitting parameters and are calculated: ,

[0042] According to the above formula, the slope and of the fitted straight line can be obtained, and the final fitted straight line model is .

[0043] Perform least squares fitting on the reference point set after RANSAC processing to obtain the left and right lane lines respectively. As shown in Figure 2 , take 10 points symmetrically and at equal intervals on the obtained left and right lane lines, and denote them as , respectively, where . Calculate the middle point according to the selected points, and the calculation formula is as follows: ,

[0044] Finally, perform least squares fitting on the middle point set to extract the navigation line.

[0045] Preferably, the step S4 is based on a pedestrian ranging algorithm with four-zone depth comparison. When moving forward along the direction of the navigation line extracted in S3, the binocular vision ranging principle is used to have two cameras observe a scene simultaneously. Based on the human eye model, the binocular camera uses the baseline distance between the two cameras (i.e., the distance between the two cameras) and the disparity between the images captured by the two cameras (i.e., the difference between the images captured by the two cameras) to calculate the depth information of the object through triangulation to achieve ranging. During detection, the detection bounding box is divided into 2×2, that is, 4 zones, and the minimum depth is calculated for each zone and compared:

[0046] S4-1: First, the YOLOv5 convolutional neural network is used to detect pedestrians in the images of the camera. The coordinate information of the detected pedestrian bounding box is ( ), and the pixel coordinates are restored according to the size of the image ( ), ( ), which respectively represent the pixel coordinates of the center point of the detection bounding box ( ), and the width and height of the inspection bounding box;

[0047] S4-2: Then, a depth map is obtained based on binocular vision from the left and right pictures. The detection box is divided into four zones: upper, lower, left, and right, and depth statistics are performed for each zone. For each zone, the average depths from the upper center point to the center point of the zone, from the bottom center point to the center line point of the zone, from the left center point to the center point of the zone, and from the right center point to the center point of the zone are respectively which respectively represent the four zones;

[0048] S4-3: Compare the average depths extracted on the left and right sides of the center point, take the smaller value of the two to get , sort all the depth values on the smaller side, and perform the same processing on the depth values at the top and bottom to get and sort them;

[0049] S4-4: Finally, compare with , select the 5 smallest depth values on the side with the smaller value and take the average as the approximate minimum depth of this zone; Finally, sort the approximate minimum depths of the four zones to get , and select the smallest depth as the depth of the pedestrian .

[0050] Preferably, in step S5, the depth range is defined as according to the pedestrian depth [Dperson - 0.2, Dperson + 0.2] The distance of the pedestrian is then obtained. The calculation formula is as follows and specifically includes: , , , distance = [Dperson - 0.2, Dperson + 0.2] ,

[0051] Figure 2 For the extraction of the navigation line in different orchard scenarios, the red line represents the left tree row line, the green represents the right tree row line, the white line represents the navigation line, and the blue dots represent the reference points for row line fitting. The experiment on the extraction of the navigation line for pictures in different orchard scenarios is shown as follows. From Figure 2 the results, it can be seen that in different orchard scenarios, the detection results of the tree trunks are relatively correct. After removing the outliers and fitting the row lines, the extracted navigation line coincides with the actual scenario. The experimental results demonstrate the effectiveness and accuracy of the algorithm proposed in the present invention.

[0052] In order to analyze the error between the navigation line extracted by the algorithm of the present invention and the manually marked navigation line, the present invention randomly selects 100 pictures as the test set in two different scenarios respectively, and conducts analysis according to the indicators in the previous section to obtain the statistical results of the angular deviation, as Figure 3 shown, and calculates the average angular deviation and the angular standard deviation. Analyzing the pictures in the test set, the statistical results of the angular deviation are obtained. From Figure 3 it can be seen that the maximum angular deviation in Scenario 1 is 7.5°, and the maximum angular deviation in Scenario 2 is 18°. Most of the angular deviations are lower than 6°. The corresponding results show that the algorithm of the present invention can meet the accuracy requirements of indoor navigation.

[0053] Through the ranging experiment using the ZED2i binocular camera, the camera is fixed on a movable device. In the simulated environment, the movable device moves forward at a constant speed, and there is a pedestrian in front of the camera's field of view. The pedestrian remains stationary at a fixed position, as Figure 4 shown in Figures 5 and 6. The results of the algorithm proposed by the present invention at 1 m, 3 m, and 8 m are visualized. The three images in each row respectively represent the ranging results in three different pedestrian postures, where the ranging unit is meters (m).

[0054] The experimental results show that within the optimal range of the binocular camera, at a distance of 1 m, when the pedestrian is in three different postures, the distances measured by the algorithm are 1.0 m, 1.1 m, and 1.2 m respectively; at a distance of 3 m, the distances are 2.7 m, 2.9 m, and 2.8 m respectively, with the maximum distance error being about 5 cm and the average error being 1.5 cm, indicating that the algorithm has high precision and verifying the accuracy of the algorithm. When exceeding the optimal range of the camera, at a distance of 8 m, in different postures, the distances measured by the algorithm are 8.1 m, 8.0 m, and 8.2 m respectively, with the maximum error being about 8 cm and the error being relatively large. The main reason is that the farther the distance from the binocular camera, the fewer the effective depth points, resulting in inaccurate distance measurement. However, in the actual scenario, the general effective distance for obstacle detection is between 1 m and 6 m. Therefore, the method of the present invention also meets the requirements for obstacle distance measurement.

[0055] As shown in Table 1, in Scenario 1, the average angular deviation of the extracted navigation line is 1.5°, and the standard deviation of the angular deviation is 1.4°, with a small error; for Scenario 2, the average angular deviation is 2.9°, and the standard deviation of the angular deviation is 2.7°, and the error is doubled compared to Scenario 1. The reason is that the perspectives of the pictures collected in Scenario 2 are variable, making the algorithm less stable. However, the error also meets the accuracy requirements for navigation line extraction, that is, the method of the present invention can adapt to the extraction of navigation lines in both scenarios.

[0056] Table 1 Analysis of error indicators for navigation line extraction in different environments Scene Number of images Average angular deviation / (°) Standard deviation of angular deviation / (°) Orchard scene 1 100 1.5 1.4 Wasteland scene 2 100 2.9 2.7 。

Claims

1. A method for unmanned system navigation line perception and obstacle ranging in a weak texture scene, characterized in that, The method includes the following steps: S1: Use the YOLOv5 detection model to perform real-time detection on the input image, obtain the coordinate information of all reference bounding boxes in the image, and extract the initial reference point set based on the midpoint at the bottom of the reference detection box; S2: Perform outlier detection and elimination on the initial reference point set obtained in step S1: Based on the RANSAC algorithm, for different detection situations, distinguish inliers and outliers, identify and eliminate the outliers deviating from the normal distribution in the point set, and obtain the reference point set; S3: Use the least squares fitting algorithm to perform least squares fitting on the reference point set obtained in step S2, respectively obtain the left and right lane lines, and calculate and extract the final navigation line again through least squares fitting based on the geometric relationship of the fitted left and right lane lines; S4: Obtain the target scene with reference to the final navigation line in step S3, divide the target scene into regions, calculate the minimum depth value of each region, compare the minimum depth values of each region, and determine the approximate minimum depth of the pedestrian; S5: Use the minimum depth obtained in S4 to construct a depth range model, obtain the depth range of the pedestrian, segment the depth image, calculate the depth mean value within the depth range of the pedestrian, and obtain the distance information of the pedestrian according to the depth mean value.

2. The method for unmanned system navigation line perception and obstacle ranging in a weak texture scenario according to claim 1, wherein The acquisition of the coordinate information of all reference bounding boxes in the image in step S1 is to extract the coordinate information of the reference bounding boxes through the geometric analysis of the target detection framework( , , , ), where,( , ) are the center point coordinates of the reference detection box, , are the width and height of the reference detection box respectively. The specific method is as follows: S1-1: Based on the geometric features of the root point of the input image, define the midpoint at the bottom of the detection bounding box as the reference point for row line fitting ( , ), and the calculation formula is: ， S1-2: Input the image to be measured into the YOLOv5 detection model, output the image detection bounding box set through feature extraction, extract the midpoint coordinates at the bottom of each detection box, obtain the initial reference points, and get the initial reference point set.

3. The method for unmanned system navigation line perception and obstacle ranging in a weak texture scene according to claim 2, wherein The specific method of step S2 is as follows: S2-1: Based on the initial reference point set obtained in step S1, and arrange the reference point set in ascending order according to the pixel coordinates; calculate the number of the reference point set ; if < 4, perform outlier detection according to the threshold method, traverse the reference point set to calculate the , difference between the current point and the reference point. If , and , then regard this point as an outlier; if > 4, then use the RANSAC algorithm to remove the outliers to obtain the finally optimized reference point set; where represents the difference in pixel coordinates on the x-axis between the current reference point and the previous reference point, represents the preset threshold for judging whether it is abnormal, represents the difference in pixel coordinates on the y-axis between the current reference point and the previous reference point, represents the preset threshold for judging whether it is abnormal; The calculation steps for RANSAC to obtain inliers in S2-2 are as follows: Set parameters for the initial data point set: predefined maximum number of iterations and error tolerance threshold , and initialize the optimal model parameters; Randomly select initial inliers: Randomly select some data points from the data points in the initial data point set as initial inliers; Fitting calculation: Generate a hypothesis model according to the selected initial inliers, calculate the fitting error between all data points in the initial data point set and the current hypothesis model, and mark the points with an error less than the threshold t as the current inliers; Update the best model: If the number of current inliers exceeds the historical best record and the model error is lower, update the optimal model parameters, the best error threshold, and the inlier set; Repeated iteration: Repeat the above steps until the preset number of iterations is reached and terminate the calculation Final inlier set: The finally obtained inlier set is the optimized reference point set.

4. The method for unmanned system navigation line perception and obstacle ranging in a weak texture scenario according to claim 3, wherein, The specific method of step S3 is as follows: S3-1: For the left and right lane line models on the two-dimensional plane, they are represented by the following formula: ， Among them, is the slope of the line, is the intercept of the line, , represent the abscissa and ordinate of the fitted line in the reference system. For the given data point , the least squares problem is described as minimizing the sum of squared residuals RSS: ， Among them, is the number of data points. By taking the derivative of RSS and setting the derivative to 0, the optimal row line fitting parameters and are estimated as follows: ， Obtain the slope of the fitted line according to the above formula and , and the final fitted line model is ; S3-2: After performing least squares fitting based on the reference point set processed by RANSAC, after obtaining the left and right lane lines respectively, take 10 points symmetrically and at equal intervals on the obtained left and right lane lines, and denote them as , , where , calculate the middle point according to the selected points. The calculation formula is as follows: ， Finally, perform least squares fitting on the intermediate point set to extract the navigation line.

5. The method for unmanned system navigation line perception and obstacle ranging in a weak texture scenario according to claim 4, characterized in that, Step S4 is a pedestrian ranging algorithm based on four-zone depth comparison. Use the binocular vision ranging principle to observe a scene with two cameras simultaneously, calculate the depth information of the object by comparing the differences between the images captured by the two cameras, and then calculate the depth of the object through triangulation to achieve ranging. When detecting, divide the detection bounding box into 2*2, that is, 4 zones, calculate the minimum depth for each zone, and make a comparison. The specific method is as follows: S4-1: First, use the YOLOv5 convolutional neural network to detect the image captured by the camera, and obtain the coordinate information of the pedestrian detection box as , and restore the pixel coordinates according to the size of the image ( ), ([[]] ) respectively represent the center point pixel coordinates ( ) of the detection bounding box, and the width of the inspection bounding box and height ;​ S4-2: Then, a depth map is obtained based on binocular vision from the left and right images. The detection box is divided into four upper, lower, left, and right partitions, and depth statistics are performed for each partition. For each partition, the average depths from the upper center point to the center point of the region, from the bottom center point to the center line point of the region, from the left center point to the center point of the region, and from the right center point to the center point of the region are represent the four regions respectively; S4-3: Depth means extracted from the left and right sides of the center point Compare and take the smaller value of the two , sort all the depth values on the smaller side, that is For the operation, the depth values at the top and bottom are processed in the same way to get And sort them; S4-4: Finally, compare with , select the 5 smallest depth values on the side with the smaller value and take the average as the approximate minimum depth of this area ; Finally, sort the approximate minimum depths of the four areas to get , and select the smallest depth as the depth of the pedestrian .

6. The method for unmanned system navigation line perception and obstacle ranging in a weak texture scene according to claim 5, characterized in that The specific method of step S5 is to obtain the pedestrian depth obtained in S4 Define the depth range as -0.2, +0.2], and obtain the distance of the pedestrian , and the calculation formula is as follows: ， ， ， 。

Images