Improved ORB-SLAM2 dense point cloud construction method based on GMS elimination

Abstract

The invention discloses an improved ORB-SLAM2 dense point cloud construction method based on GMS elimination. The method comprises the following steps: establishing an image pyramid, performing grid division on each layer of image pyramid, extracting feature points, and introducing a quadtree strategy to the feature points in each grid for screening to obtain uniform feature points; secondly, introducing a GMS algorithm in a feature matching stage to eliminate mismatching; and finally, constructing a dense point cloud map according to the pose estimation and the key frame. Through the above improvement, by improving the ORB-SLAM2 algorithm, the construction function of the dense point cloud map is increased, the dense point cloud map is not only suitable for positioning, but also can be used for navigation and path planning, the practicability of the system is improved, the generated dense point cloud map is converted into the octree topological graph through the OCTOMAP conversion tool, and the method has the advantages of being high in practicability and easy to popularize. The method is applied to automatic navigation and autonomous obstacle avoidance of unmanned aerial vehicles, unmanned vehicles and ships, and the navigation and obstacle avoidance capabilities of an autonomous system in a complex environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an improved ORB-SLAM2 dense point cloud construction method based on GMS elimination. Background Art

[0002] In the current technical environment, ORB-SLAM2 is a widely used SLAM system, which combines ORB feature detectors and descriptors, as well as backend processing based on graph optimization. Although ORB-SLAM2 performs well in a variety of environments, its performance still has room for improvement in specific scenarios, such as fast motion, drastic lighting changes, or texture-poor environments.

[0003] ORB-SLAM2 is a monocular camera-based SLAM system developed by the Computer Vision Center of the Spanish National Research Council. The system uses ORB (Oriented FAST and Rotated BRIEF) features for environmental perception and map construction. The ORB feature detector combines the advantages of FAST key point detection and BRIEF descriptor, and improves the accuracy of feature matching through rotation invariance. ORB-SLAM2 adopts a three-thread architecture: the tracking thread: responsible for tracking the movement of the camera, and determining the camera pose by matching the feature points of the current frame with the previous key frame. Local mapping thread: optimizes the local map, eliminates redundant point clouds, and ensures the consistency and accuracy of the map. Closed loop detection thread: detects and confirms closed loops to reduce cumulative errors and improve global consistency.

[0004] Other solutions mainly include: DSO (Direct Sparse Odometry): DSO is a direct SLAM system that does not rely on feature points, but directly uses pixel intensity information for pose estimation. This method performs well in environments with rich textures and stable lighting conditions, but does not work well in environments with drastic lighting changes or poor textures.

[0005] LSD-SLAM (Large-Scale Direct SLAM): LSD-SLAM is also a direct SLAM system, but it improves the robustness of the system through semi-dense maps. Nevertheless, LSD-SLAM still faces challenges in dealing with fast motion and large-scale scenes.

[0006] VINS-Mono (Visual-Inertial SLAM for Monocular Cameras): VINS-Mono combines data from vision and inertial measurement units (IMUs) to improve the stability and accuracy of the system. Although this approach solves the problems of rapid motion and lighting changes to a certain extent, it increases the complexity and cost of the system.

[0007] In the current UAV system, ORB-SLAM2, as a widely used SLAM (Simultaneous Localization and Mapping) visual mapping system, has performed well in many application scenarios, but still has certain shortcomings. The main ones are: Limitations of feature point extraction and matching: ORB-SLAM2 relies on ORB feature detectors and descriptors for environmental perception. However, in environments with fast movement, drastic lighting changes, or poor textures, the accuracy of feature point extraction and matching will drop significantly, resulting in reduced positioning accuracy. High computational complexity: ORB-SLAM2 needs to process a large amount of image data, especially in high-resolution images or high-speed motion scenes, the computational complexity is high, which may lead to insufficient real-time performance. Insufficient robustness of closed-loop detection: Although ORB-SLAM2 uses a closed-loop detection mechanism to reduce cumulative errors, the accuracy and robustness of closed-loop detection in complex environments, such as scenes with large dynamic changes, still need to be improved. Summary of the invention

[0008] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide an improved ORB-SLAM2 dense point cloud construction method based on GMS elimination. In view of the shortcomings of the prior art, the present invention establishes an image pyramid, meshes each layer of the image pyramid, extracts feature points, introduces a quadtree strategy to screen the feature points in each grid, and obtains uniform feature points; secondly, introduces the GMS algorithm in the feature matching stage to eliminate mismatches; finally, constructs a dense point cloud map based on pose estimation and key frames. Through the above improvements, the present invention aims to solve the bottleneck of the prior art in data processing and improve the overall performance of the drone system, especially the autonomous obstacle avoidance capability in complex environments.

[0009] The present invention provides the following technical solutions:

[0010] The present invention provides an improved ORB-SLAM2 dense point cloud construction method based on GMS elimination, comprising the following steps:

[0011] Step S1: Input RGB image and depth map, build image pyramid, grid each layer of image pyramid, and extract feature points;

[0012] Step S2: introduce the quadtree strategy to screen the feature points in each grid to obtain evenly distributed feature points, and then perform brute force matching;

[0013] Step S3: In the feature matching stage of step S2, the GMS algorithm is introduced to eliminate the feature points that are mismatched;

[0014] Step S4: estimate the pose of the key frame, construct a local dense point cloud map, receive image data and pose information, splice the local point cloud map, and obtain a global dense point cloud map.

[0015] As a preferred technical solution of the present invention, in step S2, a quadtree strategy is introduced to screen the feature points in each grid to obtain evenly distributed feature points. The specific process is as follows:

[0016] S401: Determine the initial number of nodes, and divide the image into a node according to the width and height of the image. Since an image has only one node in the initialization stage, the node is the root node of the quadtree;

[0017] S402: After the root node is split for the first time, four nodes are formed, and it is checked whether the feature point is used by multiple nodes;

[0018] If the point at the center of the image is used by four nodes at the same time, the number of feature points of the four nodes is counted;

[0019] If the number of feature points of a node is 0, delete the node;

[0020] If the number of feature points is 1, it will not be split any more and will be kept until the splitting is terminated and then processed;

[0021] If the number of feature points is greater than 1, the node can still be split;

[0022] Determine whether the total number of nodes at this time exceeds the set value. If not, continue to split;

[0023] S403: Continue splitting the nodes whose feature points are greater than 1, and repeat the operation in S202 until the nodes are no longer split or the number of nodes is greater than the required number of feature points;

[0024] S404: When the split reaches the node threshold, the feature point with the largest corner point response value is set as the representative feature point of this node by using non-maximum suppression, and other feature points are deleted to obtain evenly distributed feature points.

[0025] As a preferred technical solution of the present invention, the GMS algorithm in step S3 removes the feature points that are mismatched, specifically:

[0026] Convert motion smoothness constraints into statistics for eliminating false matches;

[0027] Select two images {I a ,I b}, there are N and M feature matches respectively, define C = {c1,c2,c3,…,c j} represents Figure I a To Figure I b The nearest neighbor matching is:

[0028]

[0029] In formula (1): c i —Pixel p i ,q j Matching;

[0030] N i —c i Neighborhood of;

[0031] S i —c i Similar neighborhood of ;

[0032] d—Euclidean distance between two points;

[0033] r1, r2—distance threshold;

[0034] From the binomial distribution estimation, we know that:

[0035]

[0036] In formula (2):

[0037] |S i |—S i The number of elements in c i sports support;

[0038] |N i |—c i The number of neighbors of ;

[0039] S i —c i Similar neighborhood of ;

[0040] t—the probability that a correct match is supported by a certain neighborhood window match;

[0041] ∈—the probability that an incorrect match is supported by a certain neighborhood window match;

[0042] Through calculation and analysis, we can get:

[0043]

[0044] In formula (3): Expected value;

[0045] The variance of

[0046] By setting the threshold, the expected value and variance can be expressed as:

[0047]

[0048] In formula (4): O is the distinguishability between correct and incorrect matches,

[0049] The above algorithm is accelerated by grid division. The image is divided into multiple small grids. ci represents a matching pair in grids Ga and Gb. Then the neighborhood and similar neighborhood of ci are redefined as:

[0050]

[0051] In formula (5): C a —Grid G a The matching pairs in ;

[0052] C ab —At the same time in grid G a and the grid G b The matching pairs in ;

[0053] The motion kernel method is used to solve the grid size problem. A 3x3 grid is set to convert the original neighborhood matching point problem into a grid corresponding matching problem.

[0054] Define the neighborhood of the grid as the surrounding 8 grids Therefore, the neighborhood and similarity neighborhood of ci are redefined as:

[0055]

[0056] In formula (6): C A —Grid G a of matching pairs;

[0057] C AB — matching pairs in grid A and grid B,

[0058] As a preferred technical solution of the present invention, it also includes: adopting a multi-scale and multi-rotation strategy to solve the scale and rotation problems in the matching process, specifically:

[0059] For the scale problem, we fix one grid image to be n×n in size, and the size of the other grid image to be (n·α)×(n·α). Then α has five possible values:

[0060] For the rotation problem, we use the different rotation directions of the rotation kernel to fix G a , for G b After rotating clockwise, 8 motion cores can be obtained.

[0061] As a preferred technical solution of the present invention, the step S4 estimates the pose of the key frame, constructs a local dense point cloud map, receives image data and pose information, and splices the local point cloud map to obtain a global dense point cloud map, specifically:

[0062] The principle of local dense point cloud map is camera pinhole imaging, and its mathematical model is:

[0063]

[0064] In formula (7): u is the horizontal axis of the pixel plane coordinate;

[0065] v—vertical axis of the pixel plane coordinate;

[0066] W—camera internal parameter matrix;

[0067] R—rotation matrix;

[0068] T i — coordinates of spatial points;

[0069] Q — translation matrix;

[0070] a—the factor of the ratio between the depth value and the actual space distance;

[0071] After completing the camera calibration, W is a known quantity. After the tracking thread performs pose estimation, R and Q are determined, and the obtained image depth information is used to obtain point cloud data:

[0072]

[0073] In formula (8): p is the point cloud data point;

[0074] q x —The offset of the camera optical axis on the x-axis of the image plane coordinate;

[0075] q y —The offset of the camera optical axis on the y-axis of the image plane coordinate;

[0076] f x —Camera focal length on the x-axis;

[0077] f y —Camera focal length on the y-axis;

[0078] h—depth value of the midpoint in the depth image;

[0079] After obtaining the point cloud data, a local point cloud map is generated, and point cloud splicing and global optimization are performed with the generated map to finally obtain a dense point cloud map.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] 1: The present invention introduces a quadtree strategy to screen feature points, thereby ensuring uniform distribution of feature points in the image, reducing redundancy of feature points, and improving the accuracy of feature point matching.

[0082] 2: The present invention introduces the GMS algorithm in the feature matching stage, and effectively eliminates the wrongly matched feature point pairs by converting the motion smoothness constraint into a statistic for eliminating wrong matches, thereby improving the reliability of matching.

[0083] 3: The present invention uses the idea of ​​grid division to accelerate the GMS algorithm, divides the image into multiple small grids, redefines the neighborhood and similar neighborhood of each matching pair, and significantly improves the efficiency of feature matching.

[0084] 4: The present invention adopts a multi-scale and multi-rotation strategy to solve the scale and rotation problems that occur during the matching process and enhance the adaptability of the algorithm in different scenarios.

[0085] 5: The present invention improves the ORB-SLAM2 algorithm and adds the function of building dense point cloud maps, making it not only suitable for positioning but also for navigation and path planning, thereby improving the practicality of the system. The generated dense point cloud map is converted into an octree topology map through the OCTOMAP conversion tool and applied to automatic navigation and autonomous obstacle avoidance of drones, unmanned vehicles and ships, thereby improving the navigation and obstacle avoidance capabilities of autonomous systems in complex environments.

[0086] 6: The improved algorithm of the present invention reduces the computational complexity when processing high-resolution images and high-speed motion scenes, and significantly improves real-time performance and efficiency, making it suitable for applications in various complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0088] Figure 1 It is a flow chart of an improved algorithm of the present invention that introduces the GMS algorithm;

[0089] Figure 2 is a flowchart of constructing a dense point cloud map of the present invention;

[0090] Figure 3 It is a schematic diagram of converting the neighborhood matching point problem into a grid corresponding matching problem of the present invention;

[0091] Figure 4 is a schematic diagram of the multi-scale and multi-rotation strategy of the present invention;

[0092] Figure 5It is a schematic diagram of the process of selecting feature points using the quadtree strategy of the present invention; DETAILED DESCRIPTION

[0093] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0094] Example 1

[0095] like Figure 1-5 As shown, the present invention provides an improved ORB-SLAM2 dense point cloud construction method based on GMS elimination, comprising the following steps:

[0096] Step S1: Input RGB image and depth map, build image pyramid, grid each layer of image pyramid, and extract feature points (FAST corner points);

[0097] Step S2: Introduce the quadtree strategy to screen the feature points in each grid, obtain evenly distributed feature points, and then perform brute force matching; the process is as follows Figure 5 As shown, the specific process is as follows:

[0098] S401: Determine the initial number of nodes, and divide the image into a node according to the width and height of the image. Here, the resolution of the image is 640×480. Since there is only one node in the initialization stage of an image, the node is the root node of the quadtree;

[0099] S402: After the root node is split for the first time, four nodes are formed, and it is checked whether the feature point is used by multiple nodes;

[0100] If the point at the center of the image is used by four nodes at the same time, the number of feature points of the four nodes is counted;

[0101] If the number of feature points of a node is 0, delete the node;

[0102] If the number of feature points is 1, it will not be split any more and will be kept until the splitting is terminated and then processed;

[0103] If the number of feature points is greater than 1, the node can still be split;

[0104] Determine whether the total number of nodes at this time exceeds the set value. If not, continue to split;

[0105] S403: Continue splitting the nodes whose feature points are greater than 1, and repeat the operation in S202 until the nodes are no longer split or the number of nodes is greater than the required number of feature points;

[0106] S404: When the split reaches the node threshold, the feature point with the largest corner point response value is set as the representative feature point of this node by using non-maximum suppression, and other feature points are deleted to obtain evenly distributed feature points.

[0107] Step S3: In the feature matching stage of step S2, the GMS algorithm is introduced to eliminate the feature points that are mismatched;

[0108] Specifically:

[0109] After the ORB-SLAM2 algorithm violently matches feature points, there will be mismatches. Therefore, the GMS matching algorithm is introduced in the feature matching link to eliminate mismatches. By converting the motion smoothness constraint into a statistic for eliminating mismatches, the probability of a large number of matches in the neighborhood of correct matches is high, while the probability of a large number of matches in the neighborhood of mismatches is low.

[0110] Suppose we select two images {I a ,I b}, there are N and M feature matches respectively, define C = {c1,c2,c3,…,c j} represents Figure I a To Figure I b The nearest neighbor matching is:

[0111]

[0112] In formula (1): c i —Pixel p i ,q j Matching;

[0113] N i —c i Neighborhood of;

[0114] S i —c i Similar neighborhood of ;

[0115] d—Euclidean distance between two points;

[0116] r1, r2—distance threshold;

[0117] From the binomial distribution estimation, we know that:

[0118]

[0119] In formula (2):

[0120] S i |—S i The number of elements in c i sports support;

[0121] N i |—ci The number of neighbors of ;

[0122] S i —c i Similar neighborhood of ;

[0123] t—the probability that a correct match is supported by a certain neighborhood window match; ∈—the probability that an incorrect match is supported by a certain neighborhood window match; through calculation and analysis, we can get:

[0124]

[0125] In formula (3): Expected value;

[0126] The variance of

[0127] By setting the threshold, the expected value and variance can be expressed as:

[0128]

[0129] In formula (4): O is the distinguishability between correct and incorrect matches,

[0130] The above algorithm is accelerated by grid division. The image is divided into multiple small grids. ci represents a matching pair in grids Ga and Gb. Then the neighborhood and similar neighborhood of ci are redefined as:

[0131]

[0132] In formula (5): C a —Grid G a The matching pairs in ;

[0133] C ab —At the same time in grid G a and the grid G b The matching pairs in ;

[0134] If the grid is small, the information of many neighborhoods is easily ignored. If the grid is too large, incorrect matching relationships will be included. Therefore, the motion kernel method is used to solve this problem:

[0135] The moving kernel method is used to solve the grid size problem, such as Figure 3 As shown in the figure, a 3x3 grid is set to transform the original problem of matching points in the neighborhood into a grid corresponding matching problem;

[0136] Define the neighborhood of the grid as the surrounding 8 grids Therefore, the neighborhood and similarity neighborhood of ci are redefined as:

[0137]

[0138] In formula (6): C A —Grid G a of matching pairs;

[0139] C AB — matching pairs in grid A and grid B,

[0140] The multi-scale and multi-rotation strategy is used to solve the scale and rotation problems in the matching process. Specifically:

[0141] For the scale problem, we fix one grid image to be n×n in size, and the size of the other grid image to be (n·α)×(n·α). Then α has five possible values:

[0142] For the rotation problem, we use the different rotation directions of the rotation kernel to fix G a , for G b After rotating clockwise, 8 motion cores can be obtained, such as Figure 4 shown.

[0143] Experimental derivation process:

[0144] 1. First, define the feature point sets of the two images as A and B, where A has N feature points and B has M feature points;

[0145] 2. For each feature point a i ∈A, find its nearest neighbor match b in B j ∈B, denoted as:

[0146] M={(a i ,b j )};

[0147] 3. For each matching pair (a i ,b j ), define its neighborhood as N(a i ) and the similar neighborhood is S(a i );

[0148] 4. Calculate the Euclidean distance d(a i ,b j ), and set a distance threshold ∈ to determine whether it is a correct match;

[0149] 5. According to the binomial distribution estimate, calculate the probability P that the correct match is supported by a certain neighborhood window match correct The probability P that the wrong match is supported by a neighborhood window match false ;

[0150] 6. Through calculation and analysis, it can be obtained that the expected value and variance of the correct match are E|X| and Var|X| respectively;

[0151] 7. Set the threshold, expressed as expected value and variance:

[0152]

[0153] Where P represents the distinguishability between correct and incorrect matches;

[0154] 8. Use the idea of ​​grid division to accelerate the above algorithm and divide the image into multiple small grids, denoted as C i ;

[0155] 9. Redefine the neighborhood and similarity neighborhood of each matching pair:

[0156] N(C i )={C k ∣C k With C i Adjacent}

[0157] S(C i )={C k ∣C k With C i similar and not adjacent};

[0158] 10. In order to solve the scale and rotation problems, a multi-scale multi-rotation strategy is adopted. For the scale problem, the size of one grid image is fixed to n×nn×n, and the size of the other grid image is (n·α)×(n·α)(n·α)×(n·α), where αα has 5 values. For the rotation problem, the different rotation directions of the rotating motion kernel are used. One direction is fixed, and the other direction is rotated clockwise to obtain 8 motion kernels;

[0159] 11. Through the above steps, false matches can be effectively eliminated and the accuracy of feature point matching can be improved.

[0160] Step S4: estimate the pose of the key frame, construct a local dense point cloud map, receive image data and pose information, and splice the local point cloud map to obtain a global dense point cloud map;

[0161] For positioning, the sparse feature map constructed by ORB-SLAM2 is sufficient, but for navigation and path planning, a dense point cloud map is required. Therefore, the ORB-SLAM2 algorithm is improved so that it can construct a dense point cloud map. The present invention adds the construction of a dense point cloud map on the basis of the previous improved algorithm. The process is as follows: Figure 2 As shown;

[0162] The construction of dense point cloud maps mainly includes three stages: input data, pose estimation, and map construction. First, input RGB images, depth maps, and feature points; second, perform pose estimation on key frames to build local dense point cloud maps; finally, receive image data and pose information, stitch local point cloud maps, and obtain a global dense point cloud map.

[0163] The principle of generating point cloud maps is camera pinhole imaging, and its mathematical model is:

[0164]

[0165] In formula (7): u is the horizontal axis of the pixel plane coordinate;

[0166] v—vertical axis of the pixel plane coordinate;

[0167] W—camera internal parameter matrix;

[0168] R—rotation matrix;

[0169] T i — coordinates of spatial points;

[0170] Q — translation matrix;

[0171] a—the factor of the ratio between the depth value and the actual space distance;

[0172] After completing the camera calibration, W is a known quantity. After the tracking thread performs pose estimation, R and Q are determined, and the obtained image depth information is used to obtain point cloud data:

[0173]

[0174] In formula (8): p is the point cloud data point;

[0175] q x —The offset of the camera optical axis on the x-axis of the image plane coordinate;

[0176] q y —The offset of the camera optical axis on the y-axis of the image plane coordinate;

[0177] f x —Camera focal length on the x-axis;

[0178] f y —Camera focal length on the y-axis;

[0179] h—depth value of the midpoint in the depth image;

[0180] After obtaining the point cloud data, a local point cloud map is generated, and point cloud splicing and global optimization are performed with the generated map to finally obtain a dense point cloud map.

[0181] Step S5: Read the point cloud data, publish the data content in the form of topics, use the OCTOMAP conversion tool to convert the published topic content into an octree topology map, and use the generated octree topology map for automatic navigation and autonomous obstacle avoidance of drones, unmanned vehicles, and ships.

[0182] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An improved ORB-SLAM2 dense point cloud construction method based on GMS elimination, characterized in that: The following steps are involved: Step S1: Input RGB image and depth map, build image pyramid, grid each layer of image pyramid, and extract feature points; Step S2: introduce the quadtree strategy to screen the feature points in each grid to obtain evenly distributed feature points, and then perform brute force matching; Step S3: In the feature matching stage of step S2, the GMS algorithm is introduced to eliminate the feature points that are mismatched; Step S4: estimate the pose of the key frame, construct a local dense point cloud map, receive image data and pose information, splice the local point cloud map, and obtain a global dense point cloud map.

2. According to the improved ORB-SLAM2 dense point cloud construction method based on GMS elimination according to claim 1, it is characterized in that: In step S2, the quadtree strategy is introduced to screen the feature points in each grid to obtain evenly distributed feature points. The specific process is as follows: S401: Determine the initial number of nodes, and divide the image into a node according to the width and height of the image. Since an image has only one node in the initialization stage, the node is the root node of the quadtree; S402: After the root node is split for the first time, four nodes are formed, and it is checked whether the feature point is used by multiple nodes; If the point at the center of the image is used by four nodes at the same time, the number of feature points of the four nodes is counted; If the number of feature points of a node is 0, delete the node; If the number of feature points is 1, it will not be split any more and will be kept until the splitting is terminated and then processed; If the number of feature points is greater than 1, the node can still be split; Determine whether the total number of nodes at this time exceeds the set value. If not, continue to split; S403: Continue splitting the nodes whose feature points are greater than 1, and repeat the operation in S202 until the nodes are no longer split or the number of nodes is greater than the required number of feature points; S404: When the split reaches the node threshold, the feature point with the largest corner point response value is set as the representative feature point of this node by using non-maximum suppression, and other feature points are deleted to obtain evenly distributed feature points.

3. The improved ORB-SLAM2 dense point cloud construction method based on GMS elimination according to claim 1 is characterized in that: In step S3, the GMS algorithm removes the feature points that are mismatched, specifically: Convert motion smoothness constraints into statistics for eliminating false matches; Select two images {I a ,I b }, there are N and M feature matches respectively, define C = {c1,c2,c3,...,c j } represents Figure I a To Figure I b The nearest neighbor matching is: In formula (1): c i —Pixel p i ,q j Matching; N i —c i Neighborhood of; S i —c i Similar neighborhood of ; d—Euclidean distance between two points; r1, r2—distance threshold; From the binomial distribution estimation, we know that: In formula (2): S i —S i The number of elements in c i sports support; N i —c i The number of neighbors of ; S i —c i Similar neighborhood of ; t—the probability that a correct match is supported by a certain neighborhood window match; ∈—the probability that an incorrect match is supported by a certain neighborhood window match; Through calculation and analysis, we can get: In formula (3): E Si —S i Expected value; V Si —S i The variance of By setting the threshold, the expected value and variance can be expressed as: In formula (4): P is the distinguishability between correct and incorrect matches, The above algorithm is accelerated by grid division. The image is divided into multiple small grids. ci represents a matching pair in grids Ga and Gb. Then the neighborhood and similar neighborhood of ci are redefined as: In formula (5): C a —Grid G a The matching pairs in ; C ab —At the same time in grid G a and the grid G b The matching pairs in ; The motion kernel method is used to solve the grid size problem. A 3x3 grid is set to convert the original neighborhood matching point problem into a grid corresponding matching problem. Define the neighborhood of the grid as the surrounding 8 grids Therefore, the neighborhood and similarity neighborhood of ci are redefined as: In formula (6): C A —Grid G a of matching pairs; C AB — matching pairs in grid A and grid B, 4. The improved ORB-SLAM2 dense point cloud construction method based on GMS elimination according to claim 3 is characterized in that: Also includes: The multi-scale and multi-rotation strategy is used to solve the scale and rotation problems in the matching process. Specifically: For the scale problem, we fix one grid image to be n×n in size, and the size of the other grid image to be (n·α)×(n·α). Then α has five possible values: For the rotation problem, we use the different rotation directions of the rotation kernel to fix G a , for G b After rotating clockwise, 8 motion cores can be obtained.

5. The improved ORB-SLAM2 dense point cloud construction method based on GMS elimination according to claim 1 is characterized in that: The step S4 estimates the pose of the key frame, constructs a local dense point cloud map, receives image data and pose information, and splices the local point cloud map to obtain a global dense point cloud map, specifically: The principle of local dense point cloud map is camera pinhole imaging, and its mathematical model is: In formula (7): u is the horizontal axis of the pixel plane coordinate; Δ—vertical axis of the pixel plane coordinate; W—camera internal parameter matrix; R—rotation matrix; T i — coordinates of spatial points; Q — translation matrix; a—the factor of the ratio between the depth value and the actual space distance; After completing the camera calibration, W is a known quantity. After the tracking thread performs pose estimation, R and Q are determined, and the obtained image depth information is used to obtain point cloud data: In formula (8): p is the point cloud data point; q x —The offset of the camera optical axis on the x-axis of the image plane coordinate; q y —The offset of the camera optical axis on the y-axis of the image plane coordinate; f x —Camera focal length on the x-axis; f y —Camera focal length on the y-axis; h—depth value of the midpoint in the depth image; After obtaining the point cloud data, a local point cloud map is generated, and point cloud splicing and global optimization are performed with the generated map to finally obtain a dense point cloud map.

Images