An autonomous navigation method for star catalog patrollers based on cross-temporal feature search

Abstract

The present invention provides a method for autonomous navigation of a star chart rover using cross-temporal feature search, comprising: step S1, encrypting and interpolating laser radar and visual camera point cloud data based on the density of input point cloud data; step S2, using a sliding window to detect cross-temporal feature smoothness, defining a smoothness c value, and extracting plane feature points and edge feature points based on the order of the smoothness c value; step S3, introducing a local subgraph and performing local odometry for high-frequency coarse positioning estimation; step S4, using an improved Levenberg-Marquette optimization method for global optimization and low-frequency fine mapping, completing low-frequency, high-precision global mapping using the interval period calculated using the local subgraph; and step S5, establishing a system cumulative error evaluation mechanism, revisiting locations where the system error entropy increase ΔI exceeds a threshold, and improving the overall pose estimation accuracy ΔE through a global loop-revisit detection mechanism. The present invention improves the estimation accuracy, robustness, and real-time performance of autonomous navigation.

Description

Technical Field

[0001] The present invention belongs to the field of relative positioning and perception mapping of rover, and specifically relates to a high-precision and high-robust multi-source fusion autonomous navigation method for rover in a complex environment with strong illumination, strong shadows and weak textures on a star map. Background Art

[0002] Manned star exploration, long-term resident star exploration, and star base construction have become important tasks for my country's future deep space exploration. Star rover exploration is the mobile exploration method with the highest scientific return rate. Carrying out exploration missions in unknown star environments requires real-time acquisition of precise position information of the rover. However, the unstructured environment of the star surface is complex and has the characteristics of strong illumination, strong shadows, and weak textures, which poses a strong challenge to the autonomous navigation capability of the rover. Rugged terrain, soil characteristics, and motion bumps affect the stability of the rover's driving. It is difficult to meet the accuracy and real-time requirements of autonomous navigation based on a single sensor detection method. The onboard computer has limited capabilities, and the multi-sensor fusion algorithm must consider lightweight design.

[0003] Achieving high-precision, robust, and real-time relative positioning and mapping for rovers in complex stellar environments has become a crucial research topic. The Mars rovers that have been successfully launched primarily rely on traditional inertial navigation systems and astronomical navigation systems, combined with ground-based teleoperation commands for semi-autonomous navigation. For autonomous navigation, they primarily rely on visual systems, which are selectively activated for a limited period of time in specific, texture-rich landscapes, for slow, autonomous navigation within a range of 15-75 meters. Positioning accuracy remains around 10% R, and they lack the more environmentally adaptable LiDAR systems. This autonomous navigation capability is unsuitable for future long-term autonomous exploration, impacting its real-time and reliability. Lunar rovers primarily rely on manned and unmanned ground-based teleoperation, and their autonomous navigation systems have been less validated. Therefore, there is an urgent need to develop a lightweight, multi-source fusion autonomous navigation algorithm that is robust and adaptable to complex stellar environments. Summary of the Invention

[0004] The purpose of the present invention is to provide a high-precision and high-robust multi-source fusion autonomous navigation method for a rover in a complex environment with strong illumination, strong shadows and weak textures on the star map, thereby improving the estimation accuracy, robustness and real-time performance of autonomous navigation.

[0005] To achieve the above object, the present invention provides a star chart rover autonomous navigation method for cross-time domain feature search, which comprises: step S1, according to the obstacle recognition resolution amm@bm, laser radar beam l, rover speed v, terrain obstacle particle size distribution The density of the input point cloud data is jointly derived, and the point cloud data of the lidar and visual camera are encrypted and interpolated; step S2, use the sliding window to detect the smoothness of the cross-time domain feature, define the smoothness c value, establish the feature detection cross-time domain sliding window, divide each scan into N parts, calculate the eigenvalue and eigenvector change of the local point cloud set in each part, and extract the plane feature points and edge feature points according to the sorting of the smoothness c value; step S3, introduce the local sub-graph and perform the local odometer to perform high-frequency coarse positioning estimation, store the data in the local sub-graph and Perform local point cloud registration based on Levenberg-Marquet. The conversion pose between different point cloud frames is provided by the inertial system to correct the influence of inter-frame motion distortion. Step S4: Use the improved Levenberg-Marquet optimization method to perform global optimization and low-frequency fine mapping, and complete low-frequency and high-precision global mapping at the interval period calculated by the local sub-map. Step S5: Establish a system cumulative error evaluation mechanism, revisit observations at locations where the system error entropy increase ΔI exceeds the threshold, and improve the overall pose estimation accuracy ΔE through a global loop closure-revisit detection mechanism.

[0006] Preferably, when executing step S1, the task requirements and parameters of obstacle recognition resolution amm@bm, patrol vehicle speed v, and sensor installation height h are such that the interpolation density of the laser radar and visual camera point clouds meets the requirements. The relationship between the task requirements and parameters is as follows:

[0007] b·tanα1-(bv·Δt)·tanα2≤(1 / 6)·a

[0008] Where a and b are the distance sensors at position b that can accurately identify obstacles of height a; v is the speed of the patrol vehicle; α1 and α2 are the imaging angles of the patrol vehicle after moving for Δt time.

[0009] Preferably, the feature smoothness detection method in step S2 includes: calculating the curvature of the five points before and after the current point and the current point, setting the flatness c value of the neighborhood point set in the judgment area, and the c value calculation formula is as follows:

[0010]

[0011] Where c is the calculated curvature; |S| is the number of feature points near the specified area; k is the point cloud of the kth frame; L is in the laser radar coordinate system; is the distance to the current point i.

[0012] Preferably, in step S2, each scan is divided into 4 parts, and each part is ranked according to the flatness c value, the 2 points with the largest curvature are used as edge corner points, the 4 points with the smallest curvature are used as plane feature points, and the three unstable types of features are excluded; the three unstable types of features include feature points whose angles between the surface feature normal and the laser ray are less than 5 degrees, fault feature points that are occluded as the viewing angle changes, and feature points whose pitch angles in the point cloud coordinate system are higher than the set threshold.

[0013] Preferably, the execution of the local odometer for high-frequency coarse positioning in step S3 includes the following steps: projecting the point clouds in the two scanning frames to the same time, discarding the point clouds beyond 50m and the point clouds with a minimum spacing greater than 30cm, performing local point cloud registration of the Levenberg-Marquet method, performing matching between the two scanning frames based on the correspondence between the edge corner feature points and the plane feature points, and calculating the distance between the current point and the feature.

[0014] Preferably, the edge corner feature inter-frame matching calculation is performed to construct an optimization equation for the edge corner point:

[0015]

[0016] Among them, d Γ is the shortest distance from feature point i to line jl; L is in the laser radar coordinate system; is the reprojected coordinate of the feature point i in the k+1th frame point cloud; is the coordinate of the point j closest to the feature point i in the k-th frame point cloud; is the coordinate of the point l closest to the feature point i in the adjacent scan line of the feature point j in the k-th frame point cloud;

[0017] The plane point feature inter-frame matching calculation is used to construct the optimization equation of the plane point:

[0018]

[0019] Among them, d H is the shortest distance from feature point i to plane jlm; L is in the laser radar coordinate system; is the reprojected coordinate of the feature point i in the k+1th frame point cloud; is the coordinate of the point j closest to the feature point i in the k-th frame point cloud; is the coordinate of the point l closest to the feature point i in the scan line where the feature point j is located in the k-th frame point cloud; is the coordinate of the point m closest to feature point i in the adjacent scan line of feature point j in the k-th frame point cloud.

[0020] Preferably, the laser radar inter-frame motion is assumed to be a uniform motion model, and d Γ with d HAfterwards, it is used as a constraint to solve the inter-frame conversion matrix, and the pose is solved using the Levenberg-Marquet optimization method, which is expressed as follows:

[0021]

[0022]

[0023]

[0024] Where t is the current timestamp; is the laser radar at time t relative to t k+1 changes in position and posture; is the pose transformation matrix of the i-th feature point in the k-th frame; Optimize constraints for edge corners; Optimize constraints for planar points.

[0025] Preferably, the improved Levenberg-Marquette optimization method is used in step S4 to perform global optimization and low-frequency fine mapping, match the point cloud projected to the initial scan node with the map point cloud, and optimize the global pose, including the following steps:

[0026]

[0027] in, is the updated pose in the global map at time k (low frequency); is the pose data (high frequency) during the [k+1, k+2] lidar scan frame; is the updated pose in the global map at time k+1; features consistent with the local point cloud attributes are screened globally, and the point-to-line and point-to-plane distances are calculated again; based on the point-to-line and point-to-plane distances, the Levenberg-Marquette optimization algorithm is iterated again. After first executing the Levenberg-Marquette optimization twice, points that are significantly different from the temporary results are discarded, and the iteration is continued until convergence.

[0028] Preferably, the method for establishing the system cumulative error evaluation and revisit mechanism in step S5 is as follows: the state of the patroller at time k is represented by the mean value X R (k) and variance P RR (k) Description, P RR (k) is the uncertainty of the current state estimation, using Fisher mutual information ΔI(X R (k+1),X R ) describes the entropy increase of the state estimation from time k to time k+1, and its expression is as follows:

[0029] ΔI(X R (k+1),XR )=H(X R (k+1|u k ))-H(X R (k))

[0030] Among them, H(X R (k) is the entropy increase function at time k; H(X R (k+1|u k )) is the input u at time k+1 k The entropy increase function under the condition of ΔI(X R (k+1),X R ) is the Fisher mutual information.

[0031] Preferably, for an n-dimensional vector, the corresponding relationship between the entropy increase H(x) and the logarithm of the determinant of its variance matrix is ​​as follows:

[0032]

[0033] Where det(P) is the determinant of the state variance matrix, and H(x) is the entropy increase function;

[0034] The Fisher mutual information can be simplified as:

[0035]

[0036] Where det(*) is the determinant to be solved; ΔI(*) is the Fisher mutual information expression; P is the state variance matrix; X R (k) is the state of the patroller at time k; X R (k+1) is the state of the patroller at time k+1; when the patroller travels to a certain location, the growth rate of Fisher mutual information ΔI is The information gain can be recorded dynamically; when Δt is within a period of time: Mark the key frame information at that location and determine whether to revisit the location.

[0037] In summary, compared with the existing technology, the present invention considers the strong illumination, strong shadows, weak texture, unstructured and complex environment of the star table, the accuracy and real-time requirements of the rover navigation, the limited computing power of the onboard computer, and engineering practices, and proposes a cross-temporal feature search method for the star table rover autonomous navigation, which has at least one of the following advantages:

[0038] (1) Through the multi-source fusion navigation method of laser radar, visual camera, and inertial group system, the obstacle recognition resolution and environmental perception stability are improved;

[0039] (2) Using sliding windows to detect cross-temporal feature smoothness greatly improves the real-time performance of feature extraction, feature detection efficiency, and feature matching speed;

[0040] (3) By combining multi-threaded high-frequency coarse positioning with low-frequency fine mapping and using the improved Levenberg-Marquet method for global optimization, the computational complexity is reduced and the accuracy and real-time performance of autonomous navigation are improved;

[0041] (4) A system cumulative error evaluation mechanism was established, which significantly reduced the system cumulative error by judging the revisit observations and improved the accuracy of pose estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of the autonomous navigation method for a star chart patroller provided by the present invention;

[0043] Figure 2 This is a comparison chart of the positioning accuracy between the IMU pre-integration navigation method and the fusion navigation method provided by the present invention;

[0044] Figure 3 This is a result diagram of the position error curve of the autonomous navigation of the star chart rover provided by the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the appended Figure 1 ~Attachment Figure 3 , the technical solutions, structural features, objectives achieved and effects in the embodiments of the present invention are described in detail.

[0046] It should be noted that, in the present invention, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only the elements explicitly listed, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0047] The present invention provides a star chart rover autonomous navigation method for cross-time domain feature search, such as Figure 1 Shown, including:

[0048] Step S1: According to the obstacle recognition resolution amm@bm, laser radar beam l, patrol speed v, and terrain obstacle particle size distribution The density of the input point cloud data is jointly derived, and the point cloud data of the lidar and visual camera are encrypted and interpolated;

[0049] Step S2: Use a sliding window to perform cross-temporal feature smoothness detection, define a smoothness c value, establish a cross-temporal sliding window for feature detection, divide each scan into N parts (in this embodiment, the scan is divided into 4 parts), calculate the eigenvalues ​​and eigenvector changes of the local point cloud set in each part, and extract plane feature points and edge feature points according to the order of the smoothness c value;

[0050] Step S3: Introduce a local sub-graph and perform local odometry for high-frequency coarse positioning estimation. Store data in the local sub-graph and perform local point cloud registration based on Levenberg-Marquette. The transformed poses between different point cloud frames are provided by the inertial group system to correct the influence of inter-frame motion distortion.

[0051] Step S4: Use the improved Levenberg-Marquette optimization method to perform global optimization and low-frequency fine mapping, completing low-frequency and high-precision global mapping at the interval period calculated for the local subgraphs. To reduce computing power, points with large differences from the temporary results are discarded after the first optimization, and the optimization is continued backward until iterative convergence.

[0052] Step S5: Establish a system cumulative error evaluation mechanism, revisit observations at locations where the system error entropy increase ΔI exceeds the threshold, and improve the overall pose estimation accuracy ΔE through a global loop closure-revisit detection mechanism.

[0053] When executing step S1, since at least six points are usually required to confirm obstacles in engineering judgment, the task requirements and parameters of obstacle recognition resolution amm@bm, patrol vehicle speed v, and sensor installation height h are used to ensure that the interpolation density of the lidar and visual camera point clouds meets the requirements. The relationship between the task requirements and parameters is as follows:

[0054] b·tanα1-(bv·Δt)·tanα2≤(1 / 6)·a

[0055] Where a and b are the distance sensors at position b that can accurately identify obstacles of height a; v is the speed of the patrol vehicle; α1 and α2 are the imaging angles of the patrol vehicle after moving for Δt time.

[0056] Furthermore, the feature smoothness detection method described in step S2 is as follows: calculate the curvature of the five points before and after the current point and the current point, set the flatness c value of the neighborhood point set in the judgment area, and the c value calculation formula is as follows:

[0057]

[0058] Where c is the calculated curvature; |S| is the number of feature points near the specified area; k is the point cloud of the kth frame; L is in the laser radar coordinate system; is the distance to the current i-th point;

[0059] To ensure matching uniformity, each scan is divided into 4 parts, and each part is ranked according to the flatness c value. The 2 points with the largest curvature are used as edge corner points, and the 4 points with the smallest curvature are used as plane feature points. The three types of unstable features are excluded. The three types of unstable features include feature points where the angle between the surface feature normal and the laser ray is less than 5 degrees, fault feature points that are occluded with changes in viewing angle, and feature points with a pitch angle higher than a set threshold in the point cloud coordinate system.

[0060] Furthermore, the high-frequency coarse positioning by performing local odometry in step S3 includes the following steps: projecting the point clouds in two scanning frames (one scanning frame is defined as 1 second) to the same moment, discarding point clouds beyond 50 meters and point clouds with a minimum spacing greater than 30 cm, performing local point cloud registration using the Levenberg-Marquet method, performing matching between the two scanning frames based on the correspondence between edge corner feature points and plane feature points, and calculating the distance between the current point and the feature;

[0061] The edge corner point is a point formed by the center line of the three-dimensional structure. To find the closest distance between a point and a line, we need to find the line closest to the current point. By matching the edge corner point features between frames (the k+1 frame and the k frame), we can construct the optimization equation for the edge corner point:

[0062]

[0063] Among them, d Γ is the shortest distance from feature point i to line jl; L is in the laser radar coordinate system; is the reprojected coordinate of the feature point i in the k+1th frame point cloud; is the coordinate of the point j closest to the feature point i in the k-th frame point cloud; is the coordinate of the point l closest to the feature point i in the adjacent scan line of the feature point j in the k-th frame point cloud;

[0064] The plane feature point matching is to find the correspondence between the two frames, that is, to solve the distance from the point to the plane, so it is necessary to find a plane with the closest corresponding distance. Through the plane point feature inter-frame matching calculation (the k+1 frame and the k frame), the optimization equation of the plane point is constructed:

[0065]

[0066] Among them, d H is the shortest distance from feature point i to plane jlm; L is in the laser radar coordinate system; is the reprojected coordinate of the feature point i in the k+1th frame point cloud; is the coordinate of the point j closest to the feature point i in the k-th frame point cloud; is the coordinate of the point l closest to the feature point i in the scan line where the feature point j is located in the k-th frame point cloud; is the coordinate of the point m closest to feature point i in the adjacent scan line of feature point j in the k-th frame point cloud.

[0067] Assume the inter-frame motion of the laser radar as a uniform motion model and obtain d Γ with d H After that, we need to find the right side of the corresponding minimization. We can use it as a constraint to solve the inter-frame conversion matrix and use the Levenberg-Marquette optimization method to solve the posture. Its expression is as follows:

[0068]

[0069]

[0070]

[0071] Where t is the current timestamp; is the laser radar at time t relative to t k+1 changes in position and posture; is the pose transformation matrix of the i-th feature point in the k-th frame; Optimize constraints for edge corners; Optimize constraints for planar points.

[0072] The above constraints and target optimization equation can be described as follows:

[0073]

[0074]

[0075]

[0076] Where t is the current timestamp; is the laser radar at time t relative to t k+1 The position and posture changes of ; d is the distance set between the edge corner points and the plane points; Optimization constraints for edge corner points and plane points; To solve the Jacobian matrix; λ is the Lagrange coefficient; λ(*) is a diagonal matrix; is the differential quantity, that is is the gradient descent formula. The Levenberg-Marquet optimization method is used to solve this nonlinear optimization problem, and the inter-frame transition matrix can be obtained by iterative solution.

[0077] Furthermore, the step S4 further includes:

[0078] Considering the limited onboard computing performance, the algorithm is lightweight and designed. The improved Levenberg-Marquet optimization method is used for global optimization and low-frequency fine mapping. The point cloud projected to the initial scan node is matched with the map point cloud to optimize the global pose. The steps include the following:

[0079]

[0080] in, is the updated pose in the global map at time k (low frequency); is the pose data (high frequency) during the [k+1, k+2] lidar scan frame; is the updated pose in the global map at time k+1.

[0081] Globally, features consistent with the local point cloud attributes are selected, namely, edge corner features and plane features, and the distances from point to line and point to surface are calculated again. Based on the distances from point to line and point to surface, the Levenberg-Marquette optimization algorithm is iterated again. Computational simplification should be performed here, namely, the Levenberg-Marquette optimization is performed twice first, and then points that are significantly different from the temporary results are discarded, and the iteration is continued until convergence. Computational simplification can greatly reduce the number of iterations and the amount of calculation.

[0082] Since the concept of entropy increase is reflected in the degree of reduction in the uncertainty of an estimate, the entropy increase degree description and Fisher mutual information are used to establish an entropy increase evaluation function. Further, the method for establishing the system cumulative error evaluation and revisit mechanism described in step S5 is as follows:

[0083] The state of the patroller at time k is represented by the mean value X R (k) and variance P RR (k) Description, P RR (k) reflects the uncertainty of the current state estimation, using Fisher mutual information ΔI(X R (k+1),X R ) can describe the entropy increase of the state estimation from time k to time k+1, and its expression is as follows:

[0084] ΔI(X R (k+1),X R )=H(X R (k+1|u k ))-H(X R (k))

[0085] Among them, H(X R (k) is the entropy increase function at time k; H(X R (k+1|u k )) is the input u at time k+1k The entropy increase function under the condition of ΔI(X R (k+1),X R ) is the Fisher mutual information.

[0086] For an n-dimensional vector, the corresponding relationship between the entropy increase H(x) and the logarithm of the determinant of its variance matrix is ​​as follows:

[0087]

[0088] Among them, det(P) is the determinant of the state variance matrix, and H(x) is the entropy increase function.

[0089] Furthermore, the Fisher mutual information can be simplified as:

[0090]

[0091] Where det(*) is the determinant to be solved; ΔI(*) is the Fisher mutual information expression; P is the state variance matrix; X R (k) is the state of the patroller at time k; X R (k+1) is the state of the patroller at time k+1.

[0092] When the rover travels to a certain location, the growth rate of Fisher mutual information ΔI is The information gain can be recorded dynamically. When Δt is: The key frame information of the location is marked to determine whether to revisit the location. The system cumulative error is reduced through revisit observation, effectively improving the estimation accuracy.

[0093] In one embodiment, the IMU integral navigation method and the multi-source fusion autonomous navigation method provided by the present invention are used for navigation and positioning, and the positioning accuracy results are as follows: Figure 2 As shown in the figure, by comparison, it is found that the positioning curve of the multi-source fusion autonomous navigation method of the present invention is almost consistent with the RTK true position curve, while the positioning curve of the IMU integration navigation method has a significant offset from the RTK true position curve, which shows that the positioning accuracy of the multi-source fusion autonomous navigation method provided by the present invention is more accurate. Figure 3 The position error curve of the autonomous navigation of the star chart rover shown in the figure shows that the autonomous navigation method of the star chart rover provided by the present invention effectively reduces the accumulated error of the system and further improves the estimation accuracy.

[0094] In summary, compared with the existing technology, the cross-time domain feature search star chart rover autonomous navigation method provided by the present invention improves the obstacle recognition resolution and environmental perception stability through the multi-source fusion navigation method of laser radar, visual camera, and inertial group system; the use of sliding windows for cross-time domain feature smoothness detection greatly improves the real-time performance of feature extraction, feature detection efficiency, and feature matching speed; by combining multi-threaded high-frequency coarse positioning with low-frequency fine mapping, and using the improved Levenberg-Marquette method for global optimization, the amount of calculation is reduced and the accuracy and real-time performance of autonomous navigation are improved; a system cumulative error evaluation mechanism is established, and the system cumulative error is greatly reduced by judging revisit observations, thereby improving the accuracy of pose estimation.

[0095] 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, those skilled in the art will readily appreciate various modifications and alternatives to the present invention, such as the equation for the optical system entrance pupil boundary, the tilt and rotation of the optical system entrance pupil, and the type of optical-mechanical design software used. Therefore, the scope of protection of the present invention shall be defined by the appended claims.

Claims

1. A star chart rover autonomous navigation method for cross-temporal feature search, characterized in that: include: Step S1: According to the obstacle recognition resolution amm@bm, laser radar beam l, patrol speed v, and terrain obstacle particle size distribution The density of the input point cloud data is jointly derived, and the point cloud data of the lidar and visual camera are encrypted and interpolated; Step S2: Use a sliding window to perform cross-temporal feature smoothness detection, define the smoothness c value, establish a cross-temporal sliding window for feature detection, divide each scan into N parts, calculate the eigenvalues ​​and eigenvector changes of the local point cloud set in each part, and extract plane feature points and edge feature points according to the sorting of the smoothness c value; Step S3: Introduce a local sub-graph and perform local odometry for high-frequency coarse positioning estimation. Store data in the local sub-graph and perform local point cloud registration based on Levenberg-Marquette. The transformed poses between different point cloud frames are provided by the inertial group system to correct the influence of inter-frame motion distortion. Step S4: using the improved Levenberg-Marquet optimization method to perform global optimization and low-frequency precision mapping, and completing low-frequency high-precision global mapping at the interval period calculated by the local subgraph; Step S5: Establish a system cumulative error evaluation mechanism, revisit observations at locations where the system error entropy increase ΔI exceeds the threshold, and improve the overall pose estimation accuracy ΔE through a global loop closure-revisit detection mechanism.

2. The autonomous navigation method for a star chart rover with cross-time domain feature search according to claim 1, characterized in that: When executing step S1, the task requirements and parameters of obstacle recognition resolution amm@bm, patrol vehicle speed v, and sensor installation height h are used to ensure that the interpolation density of the lidar and visual camera point clouds meets the requirements. The relationship between the task requirements and parameters is as follows: b·tanα1-(bv·Δt)·tanα2≤(1 / 6)·a Where a and b are the distance sensors at position b that can accurately identify obstacles of height a; v is the speed of the patrol vehicle; α1 and α2 are the imaging angles of the patrol vehicle after moving for Δt time.

3. The autonomous navigation method for a star chart rover with cross-temporal feature search according to claim 1, wherein: The feature smoothness detection in step S2 includes: Calculate the curvature of the five points before and after the current point and the current point, and set the flatness c value of the neighborhood point set in the judgment area. The c value calculation formula is as follows: Where c is the calculated curvature; |S| is the number of feature points near the specified area; k is the point cloud of the kth frame; L is in the laser radar coordinate system; is the distance to the current point i.

4. The autonomous navigation method for a star chart rover with cross-time domain feature search according to claim 1, characterized in that: In step S2, each scan is divided into 4 parts, and each part is ranked according to the flatness c value. The two points with the largest curvature are used as edge corner points, and the four points with the smallest curvature are used as plane feature points. The three unstable features are excluded; the three unstable features include feature points whose angle between the surface feature normal and the laser ray is less than 5 degrees, fault feature points that are occluded as the viewing angle changes, and feature points whose pitch angle in the point cloud coordinate system is higher than the set threshold.

5. The autonomous navigation method for a star chart rover with cross-time domain feature search according to claim 1, characterized in that: The execution of the local odometer for high-frequency coarse positioning in step S3 includes the following steps: The point clouds in the two scanning frames are projected to the same moment, and the point clouds beyond 50m and the point clouds with a minimum spacing greater than 30cm are discarded. The local point cloud registration of the Levenberg-Marquet method is performed. Based on the correspondence between the edge corner feature points and the plane feature points, the matching between the two scanning frames is performed, and the distance between the current point and the feature is calculated.

6. The autonomous navigation method for a star chart rover with cross-time domain feature search according to claim 5, characterized in that: The edge corner feature inter-frame matching calculation is used to construct the optimization equation of the edge corner point: Among them, d Γ is the shortest distance from feature point i to line jl; is the reprojected coordinate of the feature point i in the k+1th frame point cloud; is the coordinate of the point j closest to the feature point i in the k-th frame point cloud; is the coordinate of the point l closest to the feature point i in the adjacent scan line of the feature point j in the k-th frame point cloud; The plane point feature inter-frame matching calculation is used to construct the optimization equation of the plane point: Among them, d H is the shortest distance from feature point i to plane jlm; is the coordinate of the point l closest to the feature point i in the scan line where the feature point j is located in the k-th frame point cloud; is the coordinate of the point m closest to feature point i in the adjacent scan line of feature point j in the k-th frame point cloud.

7. The autonomous navigation method for a star chart rover with cross-time domain feature search according to claim 6, characterized in that: Assume the inter-frame motion of the laser radar as a uniform motion model and obtain d Γ with d H Afterwards, it is used as a constraint to solve the inter-frame conversion matrix, and the pose is solved using the Levenberg-Marquet optimization method, which is expressed as follows: Where t is the current timestamp; is the laser radar at time t relative to t k+1 changes in position and posture; is the pose transformation matrix of the i-th feature point in the k-th frame; Optimize constraints for edge corners; Optimize constraints for planar points.

8. The autonomous navigation method for a star chart rover with cross-time domain feature search according to claim 6, characterized in that: The improved Levenberg-Marquette optimization method described in step S4 is used to perform global optimization and low-frequency fine mapping, match the point cloud projected to the initial scan node with the map point cloud, and optimize the global pose, including the following steps: in, is the updated pose in the global map at time k (low frequency); is the pose data (high frequency) during the [k+1, k+2] lidar scan frame; is the updated pose in the global map at time k+1; globally, features consistent with the local point cloud attributes are screened, and the point-to-line and point-to-plane distances are calculated again; based on the point-to-line and point-to-plane distances, the Levenberg-Marquette optimization algorithm is iterated again. After first executing the Levenberg-Marquette optimization twice, points that are significantly different from the temporary results are discarded, and the iteration is continued until convergence.

9. The autonomous navigation method for a star chart rover with cross-time domain feature search according to claim 1, characterized in that: The method for establishing the system cumulative error evaluation and revisit mechanism described in step S5 is as follows: The state of the patroller at time k is represented by the mean value X R (k) and variance P RR (k) Description, P RR (k) is the uncertainty of the current state estimation, using Fisher mutual information ΔI(X R (k+1),X R ) describes the entropy increase of the state estimation from time k to time k+1, and its expression is as follows: ΔI(X R (k+1),X R )=H(X R (k+1|u k ))-H(X R (k)) Among them, H(X R (k) is the entropy increase function at time k; H(X R (k+1|u k )) is the input u at time k+1 k The entropy increase function under the condition of ΔI(X R (k+1),X R ) is the Fisher mutual information.

10. The autonomous navigation method for a star chart rover with cross-time domain feature search according to claim 9, characterized in that: For an n-dimensional vector, the corresponding relationship between the entropy increase H(x) and the logarithm of the determinant of its variance matrix is ​​as follows: Where det(P) is the determinant of the state variance matrix, and H(x) is the entropy increase function; The Fisher mutual information can be simplified as: Where det(*) is the determinant to be solved; ΔI(*) is the Fisher mutual information expression; P is the state variance matrix; X R (k) is the state of the patroller at time k; X R (k+1) is the state of the patroller at time k+1; When the rover travels to a certain location, the growth rate of Fisher mutual information ΔI is The information gain can be recorded dynamically; when Mark the key frame information to determine whether to revisit the location.

Images