A method for emergency return of a star surveyor

Abstract

The present invention discloses an emergency return method for a star chart rover. The method is based on environmental exploration and rut tracking, utilizes a combination of point features and line features to achieve feature complementarity, extracts, matches, and tracks historical rut information, uses the historical rut information as a guide for the return trajectory, and simultaneously introduces it into an autonomous navigation link in combination with environmental information, thereby improving the positioning accuracy of the star chart rover and also enhancing the autonomy, flexibility, and reliability of the emergency return.

Description

Technical Field

[0001] The present invention relates to the field of autonomous navigation and autonomous planning in a star rover emergency return system, and in particular to a star rover emergency return method based on environmental exploration and rut tracking. Background Art

[0002] Star rovers perform patrol and exploration missions in unknown environments. When astronauts are in emergency situations such as disability, they do not have the ability to autonomously control the movement of the rover. The rover needs to perform autonomous tracking, autonomous planning, and autonomous decision-making based on environmental information and historical track information to achieve a fast, reliable and autonomous emergency return. The same problem is encountered when the rover needs to return for resupply or perform temporary missions. The complex environment of the star surface with strong lighting, strong shadows, and weak textures, the limited wheel-soil contact force due to microgravity, the inconsistent soil softness of the star surface, and the rugged and uneven terrain of the star surface all make it difficult to detect, identify, extract, and fit wheel tracks. At the same time, adding wheel track information as a clear artificial landmark information on the star surface to autonomous navigation can enrich the navigation information source of existing rovers and improve the corresponding navigation completeness and estimation accuracy.

[0003] How to accurately identify, extract, track, and utilize rover rutting information, and achieve emergency return based on environmental exploration and rutting tracking, has become a key research topic. All successfully launched Mars rovers operate in a semi-autonomous mode, relying primarily on ground-based teleoperation. They only engage in phased, slow autonomous driving of 15-75 meters when faced with unusual terrain or when there's a risk of getting stuck. They lack a one-button emergency return function. Lunar rovers rely primarily on ground-based teleoperation and lack an emergency return function. The Apollo missions were manned lunar exploration missions, relying entirely on astronaut judgment and identification, lacking autonomous navigation, autonomous planning, or emergency return capabilities. Summary of the Invention

[0004] The purpose of the present invention is to provide an emergency return method for a star chart rover, which improves the positioning accuracy of the star chart rover and also improves the autonomy, flexibility and reliability of the emergency return.

[0005] In order to achieve the above object, the present invention provides a method for emergency return of a star rover, comprising the following steps:

[0006] Step S1: using the improved LSD algorithm to identify and detect the overall line features of the ruts to determine the rut concentration area, select the rut position in the field of view and increase the feature weight of the area;

[0007] Step S2: Use the horizontal Sobel convolution kernel to perform edge detection on the Gaussian filtered image, and obtain the horizontal rutting feature distribution through secondary Gaussian filtering and binarization;

[0008] Step S3: Use the vertical Sobel convolution kernel to perform edge detection on the Gaussian filtered image, and obtain the vertical rutting feature distribution through vertical feature fitting and Hough transform;

[0009] Step S4: projecting the extracted rutting features onto an arbitrary vertical line using a dimensionality reduction mapping histogram, retaining only the features of the middle dense area and completing the corresponding rutting centerline fitting extraction;

[0010] Step S5, establishing a rut centerline consistency error function based on the rut centerline reprojection error and introducing it into the laser visual odometry factor graph for joint optimization;

[0011] Step S6: Based on the mapping relationship between the trajectory planning results and the mission benefits, the navigation nodes with the most coverage are automatically updated to generate an online emergency return plan.

[0012] The step S1 comprises the following steps:

[0013] After acquiring the original image data, the original image is represented using a Gaussian pyramid, and pyramid downsampling and blur processing are performed. Shi-Tomasi is used for feature point detection, and the image is subjected to adaptive histogram equalization. The LK optical flow method is then used for feature point tracking. The Ransac algorithm is used to remove outliers, and a mask is set to remove dense feature points. The points with the highest tracking frequency in the area are retained to ensure that the feature points are evenly distributed.

[0014] The improved LSD algorithm is used to detect line features, the LBD algorithm is used to describe feature lines, and the KNN algorithm is used to match feature lines to complete the tracking of feature lines in the image.

[0015] Improvements to the LSD algorithm include length suppression, broken line splicing, and near-line merging. Short and long line segments are eliminated using a length-constrained elimination criterion, with thresholds set to one-eighth and two-thirds of the smaller of the image width and height. The angle difference in the direction of the feature line vectors and the pixel distance at the endpoints are used as screening principles for near-line merging and broken line splicing, with the angle difference set as the first screening principle. After angle screening, candidate line segment groups are obtained, and the horizontal and vertical distances between the feature line segment endpoints are used as distance screening thresholds.

[0016] The feature lines use the LBD descriptor algorithm in the form of floating-point vectors. The LBD descriptors are processed and converted into binary form. The Hamming distance is used to represent the similarity between matching features. The KNN algorithm is used for feature matching. The Hamming distance and angle between the feature lines to be matched are set as the screening threshold. This completes the data association of the feature lines in the imaging plane coordinate system. After that, the corresponding spatial lines in the world coordinate system are initialized by triangulating the associated feature lines.

[0017] The spatial line is represented by a combination of Plücker coordinates and orthogonal representation. Plücker coordinates are used in the calculation of feature line reprojection residuals. The spatial line is converted from the world coordinate system to the camera coordinate system as follows:

[0018]

[0019] Among them, L c is the straight line parameter in the camera coordinate system; L w is the line parameter in the world coordinate system; n c is the direction vector of the line; v c is the plane normal vector determined by the line and the coordinate origin; H cw is the transformation matrix from the world coordinate system to the camera coordinate system; (R cw ,t cw ) is the rotation and translation parameters of the transformation matrix;

[0020] The projection model of a spatial straight line onto the image plane is as follows:

[0021]

[0022] Among them, [l1,l2,l3] T is the projection coordinate; K is the projection matrix of the line feature; n c is the line direction vector; f(·) is the Jacobian matrix;

[0023] The reprojection error model is established as follows:

[0024]

[0025] Among them, e l is the reprojection error after orthogonalization, that is, matching feature segments Endpoint to projected feature segment The algebraic geometric distance of l c is the spatial line projection model; [l1,l2,l3] T is the projection coordinate; To match the starting point coordinates of the line segment; The coordinates of the end point of the matching line segment.

[0026] The step S2 comprises the following steps:

[0027] Gray the input RGB image, then downsample it, compress the image and perform Gaussian filtering, using a larger Gaussian convolution kernel to remove noise;

[0028] Use the horizontal Sobel operator to perform edge detection on the Gaussian filtered image, extract horizontal edge features, and then perform secondary Gaussian filtering;

[0029] The threshold is set for binarization, and ORB feature extraction is performed on the features extracted after binarization to obtain the lateral rutting feature distribution.

[0030] The step S3 comprises the following steps:

[0031] Gray the input RGB image, then downsample it, compress the image and perform Gaussian filtering, using a larger Gaussian convolution kernel to remove noise;

[0032] Use the vertical Sobel operator to perform edge detection on the Gaussian filtered image, extract vertical edge features, and then perform binarization processing;

[0033] After completing the vertical feature fitting, Hough transform is used to realize vertical feature detection.

[0034] The step S4 comprises the following steps:

[0035] The feature points extracted by ORB are projected onto any vertical line and dimensionality reduction mapping is performed to obtain a dimensionality reduction mapping histogram. The dense areas of projection points on both sides are filtered out, and only the ORB feature points in the dense area of ​​projection points in the middle are retained for centerline detection. A straight line fitting is performed on the centerline part to complete the fitting extraction of the rut centerline and obtain the rut centerline detection result.

[0036] The step S5 comprises the following steps:

[0037] If the detected rut line is complete, without large curvature or wheel slip ambiguity, a trajectory model consisting of the trajectory centerline is established. The trajectory posture is defined by the distance of the trajectory relative to the center of mass of the rover and the wheel slip angle as follows:

[0038]

[0039] Where D is the distance of the trajectory relative to the center of mass of the rover; α is the slip angle of the wheel; are trajectory model parameters;

[0040] A rutting centerline consistency loss function is constructed, and candidate rutting track groups are extracted using Canny edge detection and Hough transform.

[0041] The rutting trajectory errors in the image plane and real space, as well as the wheel center constraints, are set to construct a rutting centerline consistency loss function. While selecting the optimal rutting centerline, the pose of the camera and other sensors is optimized.

[0042] In the back-end optimization, the reprojection error of the characteristic line is used to construct the line feature residual and the rutting trajectory error to correct the positioning;

[0043] A joint reprojection error model of feature points and feature lines is constructed, and a sliding window method is used to jointly optimize the marginalization residual, Imu residual, and laser odometry residual to estimate the pose of the star rover and improve the pose estimation accuracy.

[0044] The joint optimization objective function is constructed as follows:

[0045]

[0046] Among them, e marg is the marginalized residual; e imu is the Imu residual; e Lidar is the residual error of laser odometry; e point is the feature point reprojection residual; e line is the line feature reprojection residual; e trace is the rutting trajectory error.

[0047] Analyze the feature line reprojection error in the sliding window, set L as the line feature set of the observed image data in the sliding window, and the line feature residual is expressed as:

[0048]

[0049] Among them, e line is the line feature residual; is the i-th frame c in the sliding window i The observed j-th spatial line L j , ρ(s) uses the Huber loss function; after residual optimization, the SLAM system framework outputs the optimized pose result of the current frame; s is the activation parameter.

[0050] The step S6 comprises the following steps:

[0051] The comprehensive evaluation function of path planning is:

[0052] f(n)=j(n)+g(n)+h(n)

[0053] Among them, f(n) is the comprehensive priority of path node n; j(n) is the track tracking accuracy; g(n) is the cost of path node n from the starting point; h(n) is the cost of path node n from the end point.

[0054] The present invention is based on environmental exploration and rut tracking, and utilizes the combination of point features and line features to achieve feature complementarity. It extracts, matches, and tracks historical rut information, uses the historical rut information as a guide for the return trajectory, and combines it with environmental information to introduce it into the autonomous navigation link, thereby improving the positioning accuracy of the satellite rover and also improving the autonomy, flexibility, and reliability of emergency return. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 The present invention provides a flow chart of a method for the emergency return of a star rover.

[0056] Figure 2 Comparison of detection line feature results between the unimproved LSD algorithm and the improved LSD algorithm.

[0057] Figure 3 This is a schematic diagram of ORB feature point detection after preprocessing.

[0058] Figure 4 It is a diagram of the dimensionality reduction projection process.

[0059] Figure 5 is the dimensionality reduction mapping histogram.

[0060] Figure 6 This is the result diagram of the rut centerline detection.

[0061] Figure 7 It is an emergency return path planning strategy diagram. DETAILED DESCRIPTION

[0062] The following is based on Figures 1 to 7 , specifically describe the preferred embodiments of the present invention.

[0063] Future developments in orbital survey missions will focus on crewed and stationed exploration. Emergency return is essential to safeguarding the lives of astronauts and the longevity of the rover. This method, not currently implemented on existing rovers, represents an urgently needed engineering technology to improve the safety, reliability, and navigation accuracy of survey missions.

[0064] like Figure 1 As shown, the present invention provides a method for emergency return of a rover based on environmental exploration and rut tracking, comprising the following steps:

[0065] Step S1: Use the improved LSD algorithm to identify and detect the overall line features of the ruts to determine the rut concentration area, select the rut position in the field of view and increase the feature weight of the area.

[0066] After obtaining the original image data, the Gaussian pyramid is used to represent the original image, and pyramid downsampling and blur processing are performed. Shi-Tomasi is used for feature point detection, and the image is adaptively histogram equalized. The LK optical flow method is then used to track feature points. The Ransac algorithm is used to eliminate outliers, and a mask is set to remove dense feature points. The points with the highest tracking times in a certain area are retained to ensure that the feature points are evenly distributed.

[0067] Tracking line feature detection requires both real-time and robust feature extraction. To ensure real-time performance and reduce the time required for line feature detection, an improved LSD line detection algorithm is used. Line features (primarily including the track centerline, track outline, and wheel track) are detected using the improved LSD algorithm. Feature lines are then described using the LBD algorithm. Finally, the KNN algorithm is used to match feature lines and track the feature lines within the image. Improvements to the LSD algorithm include length suppression, broken line splicing, and near-line merging. A length-constrained elimination criterion is used to eliminate short and long line segments. Thresholds are set at one-eighth and two-thirds of the smaller of the image width and height, respectively, increasing the LSD computation speed by three times. The angle difference between the feature line vectors and the pixel distance between their endpoints are used as the screening criteria for near-line merging and broken line splicing. The angle difference is set as the first screening criterion. After angle screening, candidate line segments are obtained. The horizontal and vertical distances between the feature line segment endpoints are used as distance screening thresholds.

[0068] The feature lines are mapped using an LBD descriptor algorithm in floating-point vector form. The LBD descriptors are processed and converted into binary form. The Hamming distance is used to represent the similarity between matching features. The KNN algorithm is then used for feature matching, with the Hamming distance and angle between the matching feature lines set as the screening threshold. This completes the data association of the feature lines in the imaging plane coordinate system. The associated feature lines are then triangulated to initialize the corresponding spatial lines in the world coordinate system.

[0069] The spatial line is represented by a combination of Plücker coordinates and orthogonal representation. Plücker coordinates are used in the calculation of feature line reprojection residuals. The spatial line is converted from the world coordinate system to the camera coordinate system as follows:

[0070]

[0071] Among them, L c is the straight line parameter in the camera coordinate system; L w is the line parameter in the world coordinate system; n c is the direction vector of the line; v c is the plane normal vector determined by the line and the coordinate origin; H cw is the transformation matrix from the world coordinate system to the camera coordinate system; (R cw ,t cw ) are the rotation and translation parameters of the transformation matrix.

[0072] The projection model of a spatial straight line onto the image plane is as follows:

[0073]

[0074] Among them, [l1,l2,l3]T is the projection coordinate; K is the projection matrix of the line feature; n c is the line direction vector; f(·) is the Jacobian matrix.

[0075] The reprojection error model is an important error model used in subsequent optimization and is established as follows:

[0076]

[0077] Among them, e l is the reprojection error after orthogonalization, that is, matching feature segments Endpoint to projected feature segment The algebraic geometric distance of l c is the spatial line projection model; [l1,l2,l3] T is the projection coordinate; The coordinates of the starting point of the matching line segment; The coordinates of the end point of the matching line segment.

[0078] By using the unimproved LSD algorithm and the improved LSD algorithm to detect the number of line features in the rut images taken in the star catalog simulation field, the results are compared as follows.

[0079] Table 1 Comparison of the number of line features detected by the unimproved LSD algorithm and the improved LSD algorithm

[0080] Detection of the number of effective features Time spent (s) No improvement on the LSD algorithm 11072 0.997443 Improved LSD algorithm 78 0.154633

[0081] Figure 2 In the figure, the left side shows the line feature detection result of the unmodified LSD algorithm, and the right side shows the line feature detection result of the improved LSD algorithm. It can be seen that the improved LSD algorithm reduces the time consumption of line feature detection and improves the calculation speed.

[0082] Step S2: Use the horizontal Sobel convolution kernel to perform edge detection on the Gaussian filtered image, and obtain the lateral rutting feature distribution through secondary Gaussian filtering and binarization.

[0083] Considering the influence of the interstellar environment, the RGB components in the RGB image do not have obvious information. The input image is grayed, then downsampled, compressed, and Gaussian filtered. A larger Gaussian convolution kernel is used to filter out noise.

[0084] Due to the camera field of view and edge detection accuracy requirements, the Sobel operator is used to perform edge detection on the Gaussian filtered image, extract horizontal edge features, and then perform secondary Gaussian filtering.

[0085] Set the threshold for binarization and perform ORB feature extraction on the features extracted by binarization, such as Figure 3 As shown, the lateral rutting characteristic distribution is obtained.

[0086] Step S3: Use the vertical Sobel convolution kernel to perform edge detection on the Gaussian filtered image, and obtain the vertical rutting feature distribution through vertical feature fitting and Hough transform.

[0087] After preprocessing and ORB feature extraction, horizontal features can be well detected and feature points can be extracted. If the extracted feature points are used to fit the centerline, the following problems will arise: (1) There is still a small probability that some feature points will be erroneously extracted due to environmental noise; (2) Affected by factors such as lighting and shooting angle, the feature points are unevenly distributed on both sides of the centerline; (3) The feature points extracted by ORB are relatively scattered, making it difficult to control the slope of the detected centerline. Therefore, it is necessary to detect and extract the vertical ruts to assist in the detection and fitting of the centerline.

[0088] For fitting of vertical features, the image is first preprocessed using the method described in step S2 above, with the difference being that a different Sobel convolution kernel is used to detect vertical edge features, and then binarization is performed.

[0089] After completing the vertical feature fitting, Hough transform is used to realize vertical feature detection.

[0090] Step S4: Use the dimension reduction mapping histogram to project the extracted rutting features onto an arbitrary vertical line, retain only the features of the middle dense area and complete the corresponding rutting centerline fitting extraction.

[0091] Based on the prior knowledge of rutting information in the image, the rutting centerline is geometrically perpendicular to the horizontal edge and parallel to the vertical feature. Therefore, the feature points extracted by ORB are projected onto any vertical line, such as Figure 4 As shown, dimensionality reduction mapping is performed to obtain a dimensionality reduction mapping histogram.

[0092] like Figure 5 As shown in the figure, the dimensionality reduction mapping histogram shows that there are three dense areas in the distribution of projection points in each area. In order to improve the detection accuracy, the dense areas on both sides are filtered out, and only the ORB feature points in the middle dense area are retained for centerline detection. The centerline part is fitted with a straight line to complete the fitting extraction of the rut centerline, and the following is obtained: Figure 6 The rut centerline detection results are shown in the figure.

[0093] Step S5: establishing a rut centerline consistency error function based on the rut centerline reprojection error and introducing it into the laser visual odometry factor graph for joint optimization to improve the positioning accuracy of multi-source information fusion.

[0094] If the detected rut line is complete and there is no large curvature or wheel slip ambiguity, a trajectory model consisting of the trajectory centerline is established. The trajectory posture is defined by the distance of the trajectory relative to the center of mass of the rover and the wheel slip angle as follows:

[0095]

[0096] Where D is the distance of the trajectory relative to the center of mass of the rover; α is the slip angle of the wheel; are trajectory model parameters.

[0097] A rutting centerline consistency loss function was constructed, and candidate rutting track groups were extracted using Canny edge detection and Hough transform. Rutting track errors in the image plane and real space, as well as wheel center constraints, were set to construct a rutting centerline consistency loss function. While selecting the optimal rutting centerline, the pose of sensors such as cameras was optimized.

[0098] In back-end optimization, the reprojection error of characteristic lines is used to construct line feature residuals and rutting trajectory errors for positioning correction. A joint reprojection error model for characteristic points and lines is constructed, and a sliding window approach is used to jointly optimize the marginalization residuals, Imu residuals, and laser odometry residuals to estimate the rover's pose and improve pose estimation accuracy.

[0099] The joint optimization objective function is constructed as follows:

[0100]

[0101] Among them, e marg is the marginalized residual; e imu is the Imu residual; e Lidar is the residual error of laser odometry; e point is the feature point reprojection residual; e line is the line feature reprojection residual; e trace is the rutting trajectory error.

[0102] The main analysis is the feature line reprojection error in the sliding window. Set L as the line feature set of the observed image data in the sliding window. The line feature residual can be expressed as:

[0103]

[0104] Among them, e line is the line feature residual; is the i-th frame c in the sliding window i The observed j-th spatial line L j , ρ(s) uses the Huber loss function. After residual optimization, the SLAM system framework can output the optimized pose result of the current frame; s is the activation parameter.

[0105] Using Euroc MH02, 04, and 05 as test data sets, the positioning method in this paper is compared with the classic VINS positioning method. The APE error comparison is shown in the following table.

[0106] Table 2 Comparison of APE error between the positioning method RUT-SLAM in this paper and the classic VINS positioning method

[0107]

[0108] Table 2 shows that after adding line features, the APE error of the proposed positioning method is reduced by 31.7% at most and 1.989% at least compared with the VINS method, and the positioning accuracy is significantly improved.

[0109] Step S6: Based on the mapping relationship between trajectory planning results and mission benefits, try to cover as many navigation nodes as possible and automatically update them to generate a highly flexible online emergency return plan.

[0110] The star chart rover will automatically generate navigation nodes when performing a patrol mission and record rutting information, environmental information, etc. in the process. When performing an emergency return mission, astronauts can set the return node and complete the remaining energy assessment of the rover. Figure 7 As shown in the figure, the path planning strategy focuses on keeping the path between the rover and the return point as short as possible and covering as many navigation points as possible to minimize path navigation errors. The rover can use either the original path rut tracking return mode or the autonomous environment exploration and optimization return mode.

[0111] The comprehensive evaluation function of path planning is:

[0112] f(n)=j(n)+g(n)+h(n)

[0113] Among them, f(n) is the comprehensive priority of path node n; j(n) is the track tracking accuracy; g(n) is the cost of path node n from the starting point; h(n) is the cost of path node n from the end point.

[0114] This paper considers the complex environment of the star system, the uniqueness and availability of star system features, the onboard computing power, and engineering practicalities, and proposes a star system rover emergency return method based on environmental exploration and track tracking. Considering that emergency return methods using single feature detection are highly dependent on the environment, particularly in scenarios with texture loss, this paper utilizes a combination of point and line features to achieve feature complementarity, extracting and tracking track information, and integrating it into the autonomous navigation process. By comprehensively utilizing multi-source information fusion navigation, this method can improve the positioning accuracy of the star system rover and enhance the autonomy, flexibility, and reliability of the emergency return.

[0115] It should be noted that in the embodiments of the present invention, the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate the description of the embodiments. They do not indicate or imply that the devices or components referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0116] Although the present invention has been described in detail through the above preferred embodiments, it should be understood that the above description is not intended to limit the present invention. After reading the above description, various modifications and substitutions of the present invention will become apparent to those skilled in the art. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for emergency return of a star rover, characterized in that: The following steps are involved: Step S1: using the improved LSD algorithm to identify and detect the overall line features of the ruts to determine the rut concentration area, select the rut position in the field of view and increase the feature weight of the area; Step S2: Use the horizontal Sobel convolution kernel to perform edge detection on the Gaussian filtered image, and obtain the horizontal rutting feature distribution through secondary Gaussian filtering and binarization; Step S3: Use the vertical Sobel convolution kernel to perform edge detection on the Gaussian filtered image, and obtain the vertical rutting feature distribution through vertical feature fitting and Hough transform; Step S4: projecting the extracted rutting features onto an arbitrary vertical line using a dimensionality reduction mapping histogram, retaining only the features of the middle dense area and completing the corresponding rutting centerline fitting extraction; Step S5, establishing a rut centerline consistency error function based on the rut centerline reprojection error and introducing it into the laser visual odometry factor graph for joint optimization; Step S6: Based on the mapping relationship between the trajectory planning results and the mission benefits, the navigation nodes with the most coverage are automatically updated to generate an online emergency return plan.

2. The method for emergency return of a star rover according to claim 1, wherein: The step S1 comprises the following steps: After acquiring the original image data, the original image is represented using a Gaussian pyramid, and pyramid downsampling and blur processing are performed. Shi-Tomasi is used for feature point detection, and the image is subjected to adaptive histogram equalization. The LK optical flow method is then used for feature point tracking. The Ransac algorithm is used to remove outliers, and a mask is set to remove dense feature points. The points with the highest tracking frequency in the area are retained to ensure that the feature points are evenly distributed. The improved LSD algorithm is used to detect line features, the LBD algorithm is used to describe feature lines, and the KNN algorithm is used to match feature lines to complete the tracking of feature lines in the image. Improvements to the LSD algorithm include length suppression, broken line splicing, and near-line merging. A length-constrained elimination criterion is used to eliminate short and long line segments, with the thresholds set to one-eighth and two-thirds of the smaller of the image width and height. The angle difference of the feature line vector direction and the pixel distance of the endpoints are used as the screening principles for merging near lines and splicing broken lines. The angle difference is set as the first screening principle. After the angle screening, the candidate line segment group is obtained. The horizontal and vertical distances of the feature line segment endpoints are used as the distance screening threshold. The feature lines use the LBD descriptor algorithm in the form of floating-point vectors. The LBD descriptors are processed and converted into binary form. The Hamming distance is used to represent the similarity between matching features. The KNN algorithm is used for feature matching. The Hamming distance and angle between the feature lines to be matched are set as the screening threshold. This completes the data association of the feature lines in the imaging plane coordinate system. After that, the corresponding spatial lines in the world coordinate system are initialized by triangulating the associated feature lines. The spatial line is represented by a combination of Plücker coordinates and orthogonal representation. Plücker coordinates are used in the calculation of feature line reprojection residuals. The spatial line is converted from the world coordinate system to the camera coordinate system as follows: Among them, L c is the straight line parameter in the camera coordinate system; L w is the line parameter in the world coordinate system; n c is the direction vector of the line; v c is the plane normal vector determined by the line and the coordinate origin; H cw is the transformation matrix from the world coordinate system to the camera coordinate system; (R cw ,t cw ) is the rotation and translation parameters of the transformation matrix; The projection model of a spatial straight line onto the image plane is as follows: Among them, [l1,l2,l3] T is the projection coordinate; K is the projection matrix of the line feature; n c is the line direction vector; f(·) is the Jacobian matrix; The reprojection error model is established as follows: Among them, e l is the reprojection error after orthogonalization, that is, matching feature segments Endpoint to projected feature segment The algebraic geometric distance of l c is the spatial line projection model; [l1,l2,l3] T is the projection coordinate; To match the starting point coordinates of the line segment; The coordinates of the end point of the matching line segment.

3. The method for emergency return of a star rover according to claim 2, wherein: The step S2 comprises the following steps: Gray the input RGB image, then downsample it, compress the image and perform Gaussian filtering, using a larger Gaussian convolution kernel to remove noise; Use the horizontal Sobel operator to perform edge detection on the Gaussian filtered image, extract horizontal edge features, and then perform secondary Gaussian filtering; The threshold is set for binarization, and ORB feature extraction is performed on the features extracted after binarization to obtain the lateral rutting feature distribution.

4. The method for emergency return of a star rover according to claim 3, wherein: The step S3 comprises the following steps: Gray the input RGB image, then downsample it, compress the image and perform Gaussian filtering, using a larger Gaussian convolution kernel to remove noise; Use the vertical Sobel operator to perform edge detection on the Gaussian filtered image, extract vertical edge features, and then perform binarization processing; After completing the vertical feature fitting, Hough transform is used to realize vertical feature detection.

5. The method for emergency return of a star rover according to claim 4, wherein: The step S4 comprises the following steps: The feature points extracted by ORB are projected onto any vertical line and dimensionality reduction mapping is performed to obtain a dimensionality reduction mapping histogram. The dense areas of projection points on both sides are filtered out, and only the ORB feature points in the dense area of ​​projection points in the middle are retained for centerline detection. A straight line fitting is performed on the centerline part to complete the fitting extraction of the rut centerline and obtain the rut centerline detection result.

6. The method for emergency return of a star rover according to claim 5, characterized in that: The step S5 comprises the following steps: If the detected rut line is complete, without large curvature or wheel slip ambiguity, a trajectory model consisting of the trajectory centerline is established. The trajectory posture is defined by the distance of the trajectory relative to the center of mass of the rover and the wheel slip angle as follows: Where D is the distance of the trajectory relative to the center of mass of the rover; α is the slip angle of the wheel; are trajectory model parameters; A rutting centerline consistency loss function is constructed, and candidate rutting track groups are extracted using Canny edge detection and Hough transform. The rutting trajectory errors in the image plane and real space, as well as the wheel center constraints, are set to construct a rutting centerline consistency loss function. While selecting the optimal rutting centerline, the pose of the camera and other sensors is optimized. In the back-end optimization, the reprojection error of the characteristic line is used to construct the line feature residual and the rutting trajectory error to correct the positioning; A joint reprojection error model of feature points and feature lines is constructed, and a sliding window method is used to jointly optimize the marginalization residual, Imu residual, and laser odometry residual to estimate the pose of the star rover and improve the pose estimation accuracy. The joint optimization objective function is constructed as follows: Among them, e marg is the marginalized residual; e imu is the Imu residual; e Lidar is the residual error of laser odometry; e point is the feature point reprojection residual; e line is the line feature reprojection residual; e trace is the rutting trajectory error. Analyze the feature line reprojection error in the sliding window, set L as the line feature set of the observed image data in the sliding window, and the line feature residual is expressed as: Among them, e line is the line feature residual; is the i-th frame c in the sliding window i The observed j-th spatial line L j , ρ(s) uses the Huber loss function; after residual optimization, the SLAM system framework outputs the optimized pose result of the current frame; s is the activation parameter.

7. The method for emergency return of a star rover according to claim 6, wherein: The step S6 comprises the following steps: The comprehensive evaluation function of path planning is: f(n)=j(n)+g(n)+h(n) Among them, f(n) is the comprehensive priority of path node n; j(n) is the track tracking accuracy; g(n) is the cost of path node n from the starting point; h(n) is the cost of path node n from the end point.

Images