A method for removing dynamic targets in SLAM by combining vision and laser radar

Abstract

The application provides a SLAM method for removing dynamic targets by combining vision and laser radar, and the method comprises the following steps: acquiring laser scanning point clouds and images in a high dynamic environment; acquiring actual dynamic targets in the images and point clouds of the dynamic targets in the laser scanning point clouds; projecting the point clouds of the dynamic targets to the images, and removing the point clouds falling in the actual dynamic targets; and performing real-time positioning and map construction based on the removed laser scanning point clouds. After the dynamic objects are removed by the application, the SLAM system can maintain high robustness in a dynamic scene, and accurate positioning and mapping can be realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mobile robot autonomous positioning, and particularly relates to a SLAM method for removing dynamic targets by combining vision and laser radar. BACKGROUND

[0002] The simultaneous localization and mapping (SLAM) technology is one of the main technologies for current mobile robot autonomous movement. First, the robot is placed in an unknown environment, and then the robot moves from an unknown starting point, uses some repeatedly observed map features to realize self-positioning and attitude recognition, and then constructs a map according to the position of the robot, and finally constructs a consistent overall environment representation through self-motion measurement and loop detection.

[0003] Laser SLAM is one of the most active research directions in the field of SLAM in recent years, which refers to the process that a mobile robot estimates its own pose by carrying a laser radar sensor, and then perceives the surrounding situation to build a map. Laser SLAM takes laser radar point cloud as input, and its advantages are that it is relatively stable, not affected by light, high in reliability, and can directly obtain the depth information of objects in the environment, and the mapping is intuitive and high in precision. In the existing laser SLAM method, the most common one is the model-based iterative matching method, which calculates the coordinate transformation between two frames of radar data by minimizing the distance of the nearest points in the continuous laser radar scan, obtains the robot pose, such as the classic LOAM and LEGO-LOAM.

[0004] However, these laser SLAM systems and point cloud registration methods are generally based on the assumption of a static environment, i.e., there are no moving objects in the background. However, in reality, the autonomous driving scene is often dynamic, and there are a large number of movable objects such as pedestrians and vehicles in the environment. When most of the feature points fall on moving objects, the point cloud registration accuracy will decrease, affecting the accuracy of positioning and mapping. Therefore, it is necessary to improve the performance of SLAM in dynamic environments. One way to solve the above problem is to construct a map containing only static objects, i.e., to remove moving object points from the point cloud map. For example, SUMA++ based on surful representation projects the three-dimensional point cloud at the current time into a two-dimensional vertex graph and a normal graph. The vertex graph is segmented using the RangeNet++ method, and a semantic weight constraint is added during point cloud registration to realize semantic ICP. And according to the semantic label of the surface element, the consistency of the semantic label of the current frame and the world model is detected, and the moving object is removed after detection. However, such methods will fail in high dynamic environments because the removal of moving targets depends on accurate poses, and accurate poses must be obtained after removing dynamic objects. Therefore, how to find a more robust positioning and mapping method in high dynamic scenes is a key problem that needs to be solved when applying SLAM algorithms. SUMMARY

[0005] To solve the above technical problems, the present application carries out research on SLAM technology based on laser vision fusion according to the characteristics of high dynamic environment, and proposes a SLAM method for jointly removing dynamic targets by vision and lidar. This method is based on laser and vision sensors, and comprehensively uses laser and image data to overcome the shortcomings of laser sensors in terms of environmental features and lack of color information, and to make up for the shortcomings of vision sensors in terms of sensitivity to light changes or low-texture environments. By combining SLAM with deep learning, dynamic targets in the mapping and positioning process are determined, the SLAM problem in high dynamic environments is converted into a SLAM problem in static environments, and the influence of dynamic targets is effectively eliminated, greatly improving the accuracy and robustness of positioning and mapping in dynamic environments.

[0006] To achieve the above purpose, the present application provides a SLAM method for jointly removing dynamic targets by vision and lidar, comprising:

[0007] acquiring laser scanning point clouds and images in a high dynamic environment;

[0008] acquiring actual dynamic targets in the images and point clouds of dynamic targets in the laser scanning point clouds;

[0009] projecting the point clouds of the dynamic targets to the images, and removing the point clouds falling within the actual dynamic targets;

[0010] Based on the pruned laser scanning point cloud, real-time positioning and map construction are performed.

[0011] Optionally, obtaining the actual dynamic target in the image comprises:

[0012] Optionally, obtaining the dynamic target candidate region in the image comprises:

[0013] Optionally, obtaining the actual dynamic target comprises:

[0014] Optionally, obtaining the dynamic target candidate region in the image comprises:

[0015] Optionally, obtaining the actual dynamic target comprises:

[0016] Optionally, obtaining the actual dynamic target comprises:

[0017] Optionally, obtaining the ORB feature points of the static region in the image comprises:

[0018] Optionally, obtaining the actual dynamic target comprises:

[0019] Optionally, obtaining the actual dynamic target comprises:

[0020] Optionally, obtaining the actual dynamic target comprises:

[0021] Optionally, obtaining the ORB feature points of the static region in the image comprises:

[0022] Optionally, obtaining the actual dynamic target comprises:

[0023] Optionally, obtaining the actual dynamic target comprises:

[0024] Optionally, obtaining the actual dynamic target comprises:

[0025] Optionally, obtaining the actual dynamic target comprises:

[0026] Based on the ORB feature points of the static region in the image, a Hamming distance between the feature points in adjacent frames is calculated to determine the matching degree, and a feature point matching pair is obtained.

[0027] Optionally, obtaining the homography matrix of the two frames of images comprises:

[0028] Suppose that there are m pairs of matching feature points in the front and back frames of images l1 and l2, and n pairs of the matching feature points are located in the static region in both frames of images. The homography matrix is constructed for the matching feature points located in the static region in both frames of images.

[0029] The homography matrix is:

[0030] p1′=Hp1

[0031] wherein p1 and p1′ represent the feature points and matching points of the previous frame l1 and the current frame l2 respectively, H represents the transformation relationship between p1 and p1′, and is denoted as the homography matrix.

[0032] Optionally, obtaining the motion feature points comprises:

[0033] Using the homography matrix H, all the feature points in the previous frame l1 are projected to the current frame l2 to obtain the coordinates after re-projection.

[0034] Based on the coordinates after re-projection, a re-projection error of the matching feature points is calculated, wherein the re-projection error comprises a static region feature point re-projection error and a non-static region feature point re-projection error.

[0035] The average value of the re-projection error of the static region is calculated.

[0036] The feature point matching pairs of the dynamic target candidate region are traversed. When the re-projection error of the feature points of the dynamic target candidate region is greater than the average value of the re-projection error of the static region, the matching points are determined as the motion feature points, otherwise, the matching points are determined as the static feature points.

[0037] Optionally, the re-projection error is:

[0038]

[0039] wherein (x1, y1) represents the coordinates of the feature point p1 in the image l1 after re-projection to the image l2, and (x′, y′) represents the coordinates of the matching point p1′ of the feature point p1 in the previous frame of image l1 in the current frame of image l2.

[0040] The average value of the re-projection error is:

[0041]

[0042] wherein ε1, ε2, and εn These represent the reprojection errors of the 1st, 2nd, and nth static feature pairs, respectively.

[0043] Optionally, the method for projecting the point cloud of the dynamic target onto the image is as follows:

[0044] Y = P * T velo_to_cam *X

[0045] Where X represents the point cloud coordinates in the lidar coordinate system, Y represents the pixel coordinates in the camera coordinate system, and T... velo_to_cam This represents the calibration parameters of the lidar to the camera, where P represents the projection matrix from the camera to the image plane.

[0046] Compared with the prior art, the present invention has the following advantages and technical effects:

[0047] This addresses the limitation of single sensors in practical applications, which cannot meet the demands of complex, variable, and structurally unstable high-dynamic environments. It fully utilizes image and point cloud information to achieve dynamic object detection and elimination.

[0048] By removing dynamic objects using this invention, the SLAM system can maintain high robustness in dynamic scenes, achieving accurate localization and mapping. Attached Figure Description

[0049] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0050] Figure 1 This is a schematic diagram of the YOLO-v5 network structure according to an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of a SLAM method for jointly eliminating dynamic targets using both vision and lidar according to an embodiment of the present invention. Detailed Implementation

[0052] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0054] This invention proposes a SLAM method for jointly eliminating dynamic targets using both vision and lidar, comprising:

[0055] acquire laser scanning point cloud and image in high dynamic environment;

[0056] acquire actual dynamic target in image and point cloud of dynamic target in laser scanning point cloud;

[0057] project point cloud of dynamic target to image and eliminate point cloud falling in actual dynamic target;

[0058] based on eliminated laser scanning point cloud, carry out real-time positioning and map construction.

[0059] Further, acquiring actual dynamic target in image comprises:

[0060] acquire dynamic target candidate area in image based on deep learning method;

[0061] carry out discrimination on dynamic target candidate area to acquire actual dynamic target.

[0062] Further, acquiring dynamic target candidate area in image comprises:

[0063] input image into deep learning network of single-stage target detection algorithm to carry out target recognition, segment different kinds of objects to obtain target detection frame, wherein, dynamic object category is used as dynamic target candidate area.

[0064] Further, discriminating dynamic target candidate area comprises:

[0065] acquire ORB feature points of static area in image; wherein, static area is area in image except dynamic target candidate area;

[0066] carry out feature matching of ORB feature points of static area between front and back frames to acquire homography matrix of two frames of image;

[0067] based on homography matrix, transform feature points of non-static area of previous frame to current frame, calculate projection error to acquire motion feature points;

[0068] when motion feature points in dynamic target candidate area are greater than preset threshold, determine that dynamic target candidate area is actual dynamic target, otherwise, determine that dynamic target candidate area is static target.

[0069] Further, acquiring ORB feature points of static area in image comprises:

[0070] carry out histogram equalization processing on image;

[0071] extract ORB feature points from processed image;

[0072] Delete the ORB feature points in the dynamic target candidate area to obtain the ORB feature points of the static area in the image.

[0073] Further, the front and back frame feature matching of the ORB feature points comprises:

[0074] Based on the ORB feature points of the static area in the image, the Hamming distance between the feature points in adjacent frames is calculated to determine the matching degree, and the feature point matching pairs are obtained.

[0075] Further, the homography matrix of the two frames of images comprises:

[0076] Suppose that the front and back frame images l1, l2 have m pairs of matching feature points, and n pairs of feature points are located in the static area in the two frames of images. The homography matrix is constructed for the matching feature points located in the static area in the two frames of images.

[0077] Further, the motion feature points comprise:

[0078] Using the homography matrix H, all the feature points in the previous frame l1 are projected to the current frame l2 to obtain the coordinates after re-projection;

[0079] Based on the coordinates after re-projection, the re-projection error of the matching feature points is calculated; wherein the re-projection error comprises: the re-projection error of the static area feature points and the re-projection error of the non-static area feature points.

[0080] The average value of the re-projection error of the static area is calculated.

[0081] The feature point matching pairs in the non-static area are traversed, and when the re-projection error of the feature points in the non-static area is greater than the average value of the re-projection error of the static area, the matching points are determined to be dynamic feature points, otherwise, the matching points are determined to be static feature points.

[0082] The embodiment aims at the problems of high dynamic point proportion, low matching precision, insufficient trajectory precision and the like of the existing laser SLAM system method in a high dynamic environment, and designs a SLAM method for removing dynamic targets by combining vision and laser radar. Based on the calibrated laser and vision sensor system, the embodiment first performs single-stage target detection on image data, and calculates a homography model by using an ORB feature detection and matching method, and obtains dynamic targets in the image by combining projection error and target detection results. Next, the corresponding point cloud data is projected to the image, and the static points in the point cloud are screened out by using the dynamic targets in the image, and a map can be constructed based on the laser scanning points corresponding to the static objects. By using the multi-sensor fusion data, reliable detection and tracking of the moving targets in the surrounding environment are realized, the influence of the moving targets on the SLAM is removed, and the positioning precision of the laser odometry is improved.

[0083] The general technical scheme of the embodiment is:

[0084] S1. Multi-target real-time detection based on yolov5

[0085] First, based on the laser and vision sensor system that has completed spatial relationship calibration, the laser scanning points and images of the environment are obtained by laser radar and camera respectively. The obtained images are input into the deep learning network of the single-stage target detection algorithm for target recognition, and different types of objects are segmented out to obtain the target detection frame and category label, target position, wherein when the category label is a possible moving category such as vehicle, pedestrian, bicycle, etc., the dynamic object category is taken as the dynamic target candidate area.

[0086] S2. Dynamic target determination based on image frames

[0087] After obtaining the dynamic target candidate area in the image in the previous step, the real motion state of the dynamic target candidate area cannot be determined. This part uses homography constraint and target detection to realize the discrimination of the dynamic target. The specific steps are as follows:

[0088] (1) Homography constraint dynamic point discrimination: based on the input image sequence, first select the ORB feature points in the static area; match the feature points in the front and rear frames to calculate the homography matrix of the two frames; use the obtained homography matrix to transform the feature points in the non-static area of the previous frame to the current frame, calculate the projection error, and obtain the motion feature points.

[0089] (2) When a certain number of motion feature points fall within the target detection frame, it is considered that the object in the area is moving and is regarded as an actual dynamic target.

[0090] S3. Dynamic point removal based on laser vision calibration parameters

[0091] Based on the input point cloud data, Euclidean clustering is performed to obtain the point cloud that may be a dynamic target; using the calibration parameters of laser vision, the point cloud in this part is projected to the image; the point cloud located in the dynamic target in the image after projection is regarded as a dynamic point and is deleted before feature extraction.

[0092] S4. Laser SLAM based on only static objects

[0093] After removing the dynamic objects, the point cloud data containing only static objects is used for real-time positioning and map construction.

[0094] The implementation process of the embodiment will be described in detail below.

[0095] S1. Multi-target real-time detection based on yolov5

[0096] In order to obtain the dynamic object candidate area, an image frame is collected by using a camera, and a YOLO v5 algorithm based on deep learning is used to detect targets in each area of the scene. In this embodiment, "pedestrian", "car" and "bicycle" are regarded as the main targets affecting positioning in the dynamic environment, and the semantic information and position of the three targets are obtained by using the YOLO-v5 target recognition method. The network structure of YOLO-v5 is as shown in Figure 1 .

[0097] S2. Dynamic target determination based on image frame

[0098] By using the deep learning method, objects such as "pedestrian", "car" and "bicycle" which have the mobility can be classified as potential dynamic targets in advance, but the real motion state cannot be judged. Therefore, dynamic target discrimination is needed. The overall flow chart is shown in Figure 2 , and the specific steps are as follows:

[0099] Step 1: Histogram equalization. The input image collected by the visual sensor is subjected to histogram equalization processing, which increases the local contrast of adjacent frame images, realizes the enhancement of the image, makes the feature extraction more balanced, and reduces the false matching.

[0100] Step 2: Extract ORB feature points and calculate descriptors. The ORB feature matching method has fast operation speed and strong robustness, and can be applied to real-time feature detection. The ORB algorithm generally consists of two parts: feature point extraction and feature point description. The feature points are extracted quickly by using the FAST algorithm, and then the feature points are described according to the BRIEF algorithm to generate feature point descriptors.

[0101] Step 3: Temporary removal of potential dynamic target feature points. The calculation accuracy of the homography matrix plays a decisive role in the accuracy of dynamic target discrimination. When solving the homography matrix, it is necessary to ensure that the extraction and matching of features are based on the static background only. Therefore, the feature points in the potential dynamic target detection box are temporarily deleted.

[0102] Step 4: ORB algorithm feature matching. Based on the remaining feature points, the Hamming distance between the feature points in adjacent frames is calculated to determine the matching degree, and at least 4 pairs of feature point matching pairs are obtained. The specific steps are as follows:

[0103] Let the descriptors of a feature point in the adjacent image to be matched be g1 and g2, and let them be XORed, then the Hamming distance between the two feature points is calculated according to formula (1).

[0104]

[0105] A certain threshold T_1 is set, and when the value of D is less than the threshold T_1, it indicates that the feature point matching is successful.

[0106] Step 5: Homography matrix calculation. The specific steps are as follows:

[0107] Suppose that the front and rear frame images l1, l2 have m pairs of matched feature points, and n pairs of feature points in the matched pairs are located in the static region in both frames. For such matched feature points, a model is constructed according to equation (2).

[0108] p1' = Hp1# (2)

[0109] In the equation, p1 and p1' represent the feature points and matched points of the previous frame l1 and the current frame l2 respectively, and H represents the transformation relationship between p1 and p1', which is denoted as a homography matrix.

[0110] According to equation (3), it is expanded as follows:

[0111]

[0112] In the equation, (x, y) and (x', y') represent the coordinates of the matched points p1 and p1' in the previous frame l1 and the current frame l2 respectively. The optimal solution of the homography matrix H of the two frames is obtained by using RANSAC.

[0113] Step 6: Dynamic point judgment by re-projection error. The specific steps are as follows:

[0114] According to equation (4), all the feature points in the previous frame l1 are projected to l2 using the obtained homography matrix H, to obtain the coordinates after re-projection.

[0115]

[0116]

[0117] According to equation (5), the re-projection error ε of the feature pair is calculated.

[0118]

[0119] In the equation, (x1, y1) represents the coordinates of the feature point p1 in the image l1 after re-projection to the image l2, and (x', y') represents the coordinates of the matched point p1' of the feature point p1 in the image l2.

[0120] The re-projection errors of the m pairs of matched feature points are calculated, including the re-projection errors of the feature points in the static region and the re-projection errors of the feature points in the non-static region. The average re-projection error ε of the static region is calculated according to equation (6). mean .

[0121]

[0122] In the equation, ε1, ε2, and ε nResidual error of re-projection of the 1st, 2nd, n th static feature pair, respectively.

[0123] The feature point matching pairs of the remaining non-static regions are traversed, and when the re-projection error ε ≤ ε mean , the matching point is considered to be a static feature point. When the re-projection error ε > ε mean , the matching point is considered to be a dynamic feature point. The matching point refers to the matching point of the traversed non-static region, that is, the matching point for which the re-projection error is calculated in the dynamic target candidate region.

[0124] Step 7: Dynamic target determination combining target detection and dynamic feature points.

[0125] Based on the target detection result of the frame image, the number k of dynamic feature points in the potential dynamic target detection frame is counted, and a threshold T2 is set. When k is greater than T2, it is determined that the potential dynamic target is moving, and is regarded as an actual dynamic target. Otherwise, it is considered that the potential dynamic target is not moving, and is regarded as a static target.

[0126] S3. Dynamic point removal based on laser vision calibration external parameters

[0127] Step 1: Point cloud clustering. The point cloud data obtained by the laser radar is processed, and the point cloud is clustered into ground points, background points, and foreground points by using the Euclidean clustering method, wherein the foreground points are regarded as possible dynamic points in the point cloud.

[0128] Step 2: Point cloud dynamic target elimination based on external parameter matrix. First, the actual dynamic target is determined according to the above method by using the image obtained by the vision sensor. Second, the mapping relationship from the laser radar coordinate system to the camera coordinate system is obtained through the fusion correction of the laser sensor and the vision sensor. According to formula (7), the foreground points obtained by clustering the point cloud are projected to the corresponding image, and the point cloud falling within the actual dynamic target is regarded as a dynamic point and is deleted before feature matching, finally realizing accurate removal of dynamic objects in the point cloud and reducing the interference of dynamic targets on the SLAM process.

[0129] Y = P * T velo_to_cam *X#(7)

[0130] In the formula, X represents the point cloud coordinates in the laser radar coordinate system, Y represents the pixel coordinates in the camera coordinate system, T velo_to_cam represents the calibration parameters of the laser radar to the camera, including the rotation matrix and the translation matrix, and P represents the projection matrix of the camera to the image plane.

[0131] S4. Laser SLAM based on only static objects

[0132] After the dynamic objects are removed, the point cloud data containing only static objects is used for real-time positioning and map construction. This part is based on the classic LOAM work. First, the point cloud data is preprocessed to remove external points and noise. Second, the curvature of the point cloud is calculated, the edge points and plane points are extracted as feature points, and the self-motion of the laser radar is estimated by minimizing the distance between the target edge feature points and the corresponding straight line and minimizing the distance between the target plane feature points and the corresponding plane. Finally, a high-precision static point cloud map is constructed based on the odometer.

[0133] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A SLAM method for visual and laser radar combined rejection of dynamic targets, characterized in that, The method comprises the following steps: acquiring laser scanning point cloud and image in high dynamic environment; acquiring actual dynamic target in the image and point cloud of dynamic target in the laser scanning point cloud; projecting the point cloud of the dynamic target to the image and eliminating the point cloud falling in the actual dynamic target; carrying out real-time positioning and map construction based on the eliminated laser scanning point cloud; acquiring actual dynamic target in the image comprises: acquiring dynamic target candidate area in the image based on deep learning method; carrying out discrimination on the dynamic target candidate area to acquire actual dynamic target; the discrimination on the dynamic target candidate area comprises: acquiring ORB feature points of static area in the image; wherein the static area is the area in the image except the dynamic target candidate area; carrying out front and back frame feature matching on the ORB feature points of the static area to acquire homography matrix of two frames of images; based on the homography matrix, transforming feature points of the dynamic target candidate area in the previous frame to the current frame, calculating projection error to acquire motion feature points; when the motion feature points in the dynamic target candidate area are greater than a preset threshold, determining that the dynamic target candidate area is the actual dynamic target, otherwise, determining that the dynamic target candidate area is static target. 2.The SLAM method of visual and laser radar combined dynamic target rejection according to claim 1, wherein, acquiring dynamic target candidate area in the image comprises: carrying out target recognition on the image, segmenting different kinds of objects to obtain target detection frame and category label, target position, wherein dynamic object category is taken as the dynamic target candidate area; the dynamic object category is: the object whose category label includes vehicle, pedestrian and bicycle. 3.The SLAM method of visual and laser radar combined dynamic target rejection according to claim 1, wherein, acquiring ORB feature points of static area in the image comprises: carrying out histogram equalization processing on the image; extracting ORB feature points from the processed image; deleting ORB feature points in the dynamic target candidate area to acquire ORB feature points of static area in the image. 4.The SLAM method of visual and laser radar combined dynamic target rejection according to claim 1, wherein, carrying out front and back frame feature matching on the ORB feature points comprises: based on the ORB feature points of static area in the image, calculating Hamming distance between feature points in adjacent frames to determine matching degree and acquire feature point matching pair. 5.The SLAM method of visual and laser radar combined dynamic target rejection according to claim 1, wherein, acquiring homography matrix of two frames of images comprises: Set the front and back frame images , n pairs of matched feature points, wherein n pairs of matched feature points are located in the static region in both frames of images, and the homography matrix is constructed for the matched feature points located in the static region in both frames of images; the homography matrix is: in, These represent the previous frame respectively. and the current frame Feature points and matching points, H represents The transformation relationship between them is denoted as the homography matrix. 6.The SLAM method of visual and laser radar combined dynamic target rejection according to claim 1, wherein, acquiring the motion feature points comprises: Using homography matrix H, project all feature points in the previous frame to the current frame , and obtain the re-projected coordinates; based on the reprojected coordinates, calculating re-projection error of matched feature points; wherein the re-projection error includes: static area feature point re-projection error and non-static area feature point re-projection error; calculating mean value of re-projection error of static area; iterating feature point matching pair of dynamic target candidate area, when re-projection error of feature point of dynamic target candidate area is greater than the mean value of re-projection error of static area, determining that the matching point is the motion feature point, otherwise, determining that the matching point is static feature point.

7. The SLAM method of visual and laser radar combined dynamic target rejection according to claim 6, characterized in that, the re-projection error is: in,( , ) represents an image Feature points in Reprojection to image The coordinates after, ( , () indicates the previous frame image Feature points in In the current frame image middle matching point The coordinates; the mean value of re-projection error is: wherein, , respectively represent the re-projection error of the 1st, 2nd, nth static feature pair. 8.The SLAM method of visual and laser radar combined dynamic target rejection according to claim 1, wherein, the method for projecting the point cloud of the dynamic target to the image is: Wherein, X represents the point cloud coordinates in the laser radar coordinate system, Y represents the pixel coordinates in the camera coordinate system, represents the calibration parameters of the laser radar to the camera, .

Images