Visual SLAM method based on dynamic attribute cognition
By combining ORB feature point extraction, YOLO target recognition, LK optical flow method and RANSAC dynamic attribute judgment, the problem of insufficient positioning accuracy and robustness of visual SLAM systems in dynamic environments is solved, and more accurate dynamic area judgment and positioning are achieved Position estimate.
Patent Information
- Application Number
- CN202410922131.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-10
- Publication Date
- 2024-11-22
AI Technical Summary
The existing visual SLAM system has a great influence on the positioning accuracy and robustness of dynamic objects in dynamic environments, especially in unclear texture scenarios, and it is impossible to effectively judge the status of passive moving objects.
Using a method based on dynamic attribute cognition, ORB feature point extraction, YOLO target recognition, LK optical flow method and RANSAC dynamic attribute judgment are used, dynamic and static areas are judged in combination with affiliation, dynamic feature points are eliminated, and pose estimation accuracy is improved.
More accurate dynamic area judgment is achieved, the accuracy and recognition speed of the algorithm are improved, and the impact of dynamic objects on SLAM system map construction is reduced.
Smart Images

Figure CN119006589A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of computer vision and robot mapping, and in particular to a visual SLAM method based on dynamic attribute cognition. Background Art
[0002] Visual SLAM (Simultaneous Localization and Mapping) is currently widely used in the field of robotics. Robots can use visual sensors to collect information in an unknown environment to estimate their current position and build an incremental map of the current environment. This incremental map is usually represented in the form of point clouds or feature points, which contains key information about the robot's surroundings. Most of the operating conditions of this system are based on a static environment, but in most real-world scenarios, there will be dynamic objects, such as pedestrians, animals, vehicles, and other dynamic objects. If dynamic objects occupy a relatively large proportion in the field of view, it will have a serious impact on the positioning accuracy and robustness of the system.
[0003] Research on visual localization problems in dynamic scenes can be divided into two main categories: one is based on traditional multi-view geometry methods and related improved methods, and the other is a method combined with deep learning.
[0004] Based on the traditional multi-view geometry method and its related improved methods, pixels are divided into dynamic and static using two-level geometric constraints. The distance between the matching point and the polar line is used as the basis for judging the motion state of the feature point, and the FVB (Fundamental Verifi-cation Bound) constraint is used to measure the distance between the spatial point and the projection point in the image along the limit motion. The motion state of the feature point is judged based on the motion parameters of the conjunction. This algorithm requires a large amount of calculation of the motion state of the feature point, and the algorithm runs slowly. CoSLAM is used to track the reprojection of feature points in the threshold size relationship, and the motion state of the feature points is judged. Dynamic constraints are introduced into the optimization function to improve the running efficiency and robustness of the algorithm. The dense 3D optical flow of the image is roughly calculated using continuous frames, and the derived residual motion likelihood is used to detect dynamic objects, thereby improving the performance of the SLAM system in dynamic environments, but the system is less effective in scenes with unclear textures.
[0005] Among the methods combined with deep learning, there are methods that use the system to add semantic threads to add semantic labels to potential dynamic objects, obtain feature points of objects that may move through the obtained prior semantic information, and reduce dynamic feature points. However, this method only proposes prior dynamic objects and cannot judge the state of passive moving objects. Mask-RCNN and multi-view geometry methods are used to detect dynamic objects, and dynamic points are eliminated on this basis. The DynaSLAM algorithm is proposed based on the ORB-SLAM2 system. There are also methods that use SSD (Single Shot MultiBox Detector) to detect objects in images, mark the types of dynamic objects according to prior probabilities, and filter out feature points of dynamic objects. SegNet is used to obtain semantic information in the image, and on this basis, dynamic points are eliminated by combining motion consistency detection. The DS-SLAM (semantic visual SLAM toward dynamic environments) algorithm is proposed based on ORB-SLAM, but the system has poor robustness in scenes with large dynamic areas and poor segmentation of dynamic areas. Summary of the invention
[0006] The purpose of the present invention is to address the deficiencies of the prior art and provide a visual SLAM method based on dynamic attribute cognition. By combining the dynamic attributes and affiliation of the target, a more accurate determination of dynamic and static areas is achieved, and then the feature points of the dynamic area are removed to obtain a more accurate pose estimation, thereby reducing the impact of dynamic objects on the map construction of the visual SLAM system.
[0007] To achieve the above purpose, the present invention adopts the following technical solution.
[0008] A visual SLAM method based on dynamic attribute recognition includes the following steps:
[0009] Step S1, performing ORB (Oriented FAST and Rotated BRIEF) feature point extraction on the image to obtain feature points of the image;
[0010] Step S2, using YOLO (Oriented FAST and Rotated BRIEF, directional fast and rotation invariant binary) to perform target recognition on the image, detect various targets in the image, and input the two-dimensional position information of various targets into the ORB-SLAM2 (ORB-based monocular vision SLAM system 2) algorithm;
[0011] Step S3, performing affiliation determination on the passive targets around the active target to obtain the passive targets related to the active target;
[0012] Step S4, performing dynamic attribute determination using the LK optical flow method and RANSAC on the obtained active target and the passive target region having a subordinate relationship with the active target;
[0013] Step S5: Remove the feature points in the active target and passive target areas that are finally determined to be dynamic and not in the static target area, and perform subsequent operations of ORB-SLAM2 on the remaining static feature points.
[0014] Specifically, the step S1 of extracting ORB feature points from an image includes the following steps:
[0015] Step S11, extracting feature points from the current image, and when the number of detected feature points exceeds a set threshold, initialization is considered successful;
[0016] Step S12: Calculate the main direction of the ORB feature point using the grayscale centroid method, and define the moment m of the image block as:
[0017] m pq =Σ x,y∈B x p y q I(x,y),p,q={2,1} (1)
[0018] In the above formula: p is the offset of the pixel; q is the offset of the pixel; x is the horizontal coordinate of the pixel; y is the vertical coordinate of the pixel; B is the image block area; I (x, y) is the gray value of this coordinate;
[0019] Find the centroid C of the image block through the moment m:
[0020]
[0021] Connect the geometric center O and the centroid C of the image block to obtain a direction vector OC, which defines the direction of the feature point as:
[0022]
[0023] Then, ORB feature points with direction information are obtained.
[0024] Specifically, in step S2, YOLO is used to identify the target in the image, detect various targets in the image, and input the two-dimensional position information of various targets into the ORB-SLAM2 algorithm. The process is as follows:
[0025] Step S21, train YOLO using the COCO2017 dataset to obtain training weights, and use the trained YOLO to perform target recognition on the image;
[0026] Step S22: Classify the target according to the target recognition result into the following three types:
[0027] a. Static targets: objects that do not move for a long time and are difficult to move, such as walls, pillars, wall-mounted air conditioners, tables, refrigerators, cabinets, etc.;
[0028] b. Active targets: entities that can move autonomously, such as people, cats, cars, etc.
[0029] c. Passive target: an object that moves due to the action of the active target, such as a mouse, a teacup, a bicycle, etc.
[0030] Based on the three types, dynamic attributes are divided into each target. Static objects are considered static unless there are special circumstances. Active targets are divided into transient and dynamic, and passive targets are divided into static, transient and dynamic.
[0031] Among them, static refers to a state of not moving for a period of time; transient refers to a state with a probability of movement; dynamic refers to a state in motion;
[0032] Step S23: input the two-dimensional position information of the target after classification in step S22 into the ORB-SLAM2 algorithm.
[0033] Specifically, the process of determining the affiliation of the passive targets around the active target in step S3 to obtain the passive targets related to the active target is as follows:
[0034] Step S31, according to the various targets in the image detected in step S2, the active targets in the image are processed in sequence, firstly, the targets that are not human in the active targets are determined, and the Euclidean distance between the center point of the active target and the center points of the surrounding passive targets is directly used to determine whether the distance is less than a set threshold value, and when the Euclidean distance is less than the set threshold value, it is determined that there is a subordinate relationship between the two;
[0035] Step S32: When the active target is a human, additionally identify the key points of the human hand, and determine the distance between the passive target in the image and the key points of the hand. When the distance is less than the set threshold, it is determined that the passive target and the active target have a subordinate relationship and have the same dynamic properties.
[0036] Specifically, the dynamic attribute determination process in step S4 is as follows:
[0037] Step S41, dynamic attribute determination based on LK optical flow method;
[0038] Let the grayscale of a pixel at (x, y) at time t be recorded as I(x, y, t). Then for the pixel at (x, y) at time t, it moves to (x+dx, y+dy) at time t+dt, we have:
[0039] I(x+dx,y+dy,t+dt)=I(x,y,t) (4)
[0040] Based on the grayscale invariance assumption of the LK optical flow method, the grayscale at the next moment is equal to the grayscale at the previous moment, so equation (4) is further transformed into:
[0041]
[0042] In the above formula, is the gradient I of the image on the x-axis at that point x ; is the gradient I of the image on the y axis at that point y ; is the speed v of the pixel on the x-axis x ; is the velocity v on the y-axis y ;
[0043] The change of image grayscale with respect to time is recorded as I t ,get:
[0044]
[0045] Transform the above formula (6) into x , v y The overdetermined linear equation is solved using the least squares method to obtain the x and y velocities, thereby determining the dynamics at (x, y);
[0046] Step S42: dynamic attribute determination based on RANSAC;
[0047] Match the feature points of the current frame with the feature points of the reference frame to establish a correspondence between the two frames; use the RANSAC algorithm to estimate the motion model between the two frames to obtain the motion posture of the feature points;
[0048] In the process of motion estimation, RANSAC determines which matching points are inliers, i.e. points that conform to the motion model, and which are outliers, i.e. points that do not conform to the motion model, based on the error between the motion model and the matching points.
[0049] Step S43: Based on the method in step S41 and step S42, dynamic attribute determination is performed on the obtained active target and the passive target area having a subordinate relationship with the active target, and feature points in the area where they are located and are not static target areas are eliminated.
[0050] Specifically, step S5 specifically includes the following steps:
[0051] Step S51, combining YOLO, LK optical flow method, and RANSAC to determine the dynamic properties of active targets and passive target areas having a subordinate relationship therewith;
[0052] Step S52, removing feature points in the active target and passive target areas that are finally determined to be dynamic and not in the static target area;
[0053] Step S53: After removing the dynamic feature points, the static feature points are used to complete the subsequent operations of the ORB-SLAM2 algorithm.
[0054] Compared with the prior art, the present invention has the following beneficial effects:
[0055] 1. The method of the present invention uses the affiliation relationship between targets to make more accurate dynamic area determination;
[0056] 2. The method of the present invention uses a lightweight network for feature processing, which effectively improves the speed and accuracy of recognizing semantic information in images;
[0057] 3. The present invention combines the LK optical flow method to assist in judging the dynamics of the target area, which can effectively improve the accuracy of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 The present invention is a flow chart of a visual SLAM method based on dynamic attribute cognition.
[0059] Figure 2 It is the absolute trajectory error map ATE of the present invention in the fr3_walking_xyz data set.
[0060] Figure 3 This is the relative trajectory error map RTE of the present invention in the fr3_walking_xyz data set. DETAILED DESCRIPTION
[0061] In order to better reflect the advantages and innovations of the present invention and clearly demonstrate the technical solution of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments.
[0062] Example
[0063] like Figure 1 As shown, the present invention provides a visual SLAM method based on dynamic attribute cognition, comprising the following steps:
[0064] Step S1, performing ORB (Oriented FAST and Rotated BRIEF) feature point extraction on the image to obtain feature points of the image;
[0065] Step S2, using YOLO (Oriented FAST and Rotated BRIEF, directional fast and rotation invariant binary) to perform target recognition on the image, detect various targets in the image, and input the two-dimensional position information of various targets into the ORB-SLAM2 (ORB-based monocular vision SLAM system 2) algorithm;
[0066] Step S3, performing affiliation determination on the passive targets around the active target to obtain the passive targets related to the active target;
[0067] Step S4, performing dynamic attribute determination using the LK optical flow method and RANSAC (Random Sample Consensus) on the obtained active target and the passive target region having a subordinate relationship with the active target;
[0068] Step S5: Remove the feature points in the active target and passive target areas that are finally determined to be dynamic and not in the static target area, and perform subsequent operations of ORB-SLAM2 on the remaining static feature points.
[0069] Specifically, the step S1 of extracting ORB feature points from an image includes the following steps:
[0070] Step S11, extracting feature points from the current image, and when the number of detected feature points exceeds a set threshold, initialization is considered successful;
[0071] Step S12: Calculate the main direction of the ORB feature point using the grayscale centroid method, and define the moment m of the image block as:
[0072] m pq =Σ x,y∈B x p y q I(x,y),p,q={2,1} (1)
[0073] In the above formula: p is the offset of the pixel; q is the offset of the pixel; x is the horizontal coordinate of the pixel; y is the vertical coordinate of the pixel; B is the image block area; I (x, y) is the gray value of this coordinate;
[0074] Find the centroid C of the image block through the moment m:
[0075]
[0076] Connect the geometric center O and the centroid C of the image block to obtain a direction vector OC, which defines the direction of the feature point as:
[0077]
[0078] Then, ORB feature points with direction information are obtained.
[0079] Specifically, in step S2, YOLO is used to identify the target in the image, detect various targets in the image, and input the two-dimensional position information of various targets into the ORB-SLAM2 algorithm. The process is as follows:
[0080] Step S21, train YOLO using the COCO2017 dataset to obtain training weights, and use the trained YOLO to perform target recognition on the image;
[0081] In this embodiment, the YOLO system consists of five parts: input end, feature extraction network (Backbone), feature fusion network (Neck), prediction network (Prediction) and output end.
[0082] Among them, the input end adopts the Mosaic data enhancement method, and splices pictures by random scaling, cropping and arrangement to realize adaptive anchor frame calculation and picture scaling of the original picture, thereby improving the small target detection effect; in terms of Backbone, the backbone network uses CSP1_1 and CSP1_3 structures, and the Focus structure of the feature extraction part uses 32 convolution kernels to slice the picture; the Neck backbone network adopts the network structure of FPN combined with PAN, and the included CSP2_1 network structure significantly enhances the feature fusion ability of the algorithm.
[0083] Step S22: Classify the target according to the target recognition result into the following three types:
[0084] a. Static targets: objects that do not move for a long time and are difficult to move, such as walls, pillars, wall-mounted air conditioners, tables, refrigerators, cabinets, etc.;
[0085] b. Active targets: entities that can move autonomously, such as people, cats, cars, etc.
[0086] c. Passive target: an object that moves due to the action of the active target, such as a mouse, a teacup, a bicycle, etc.
[0087] Based on the three types, dynamic attributes are divided into each target. Static objects are considered static unless there are special circumstances. Active targets are divided into transient and dynamic, and passive targets are divided into static, transient and dynamic.
[0088] Among them, static refers to a state of not moving for a period of time; transient refers to a state with a probability of movement; dynamic refers to a state in motion;
[0089] Step S23: input the two-dimensional position information of the target after classification in step S22 into the ORB-SLAM2 algorithm.
[0090] Specifically, the process of determining the affiliation of the passive targets around the active target in step S3 to obtain the passive targets related to the active target is as follows:
[0091] Step S31, according to the various targets in the image detected in step S2, the active targets in the image are processed in sequence, firstly, the targets that are not human in the active targets are determined, and the Euclidean distance between the center point of the active target and the center points of the surrounding passive targets is directly used to determine whether the distance is less than a set threshold value, and when the Euclidean distance is less than the set threshold value, it is determined that there is a subordinate relationship between the two;
[0092] Step S32: When the active target is a human, additionally identify the key points of the human hand, and determine the distance between the passive target in the image and the key points of the hand. When the distance is less than the set threshold, it is determined that the passive target and the active target have a subordinate relationship and have the same dynamic properties.
[0093] Specifically, the dynamic attribute determination process in step S4 is as follows:
[0094] Step S41, dynamic attribute determination based on LK optical flow method;
[0095] Let the grayscale of a pixel at (x, y) at time t be recorded as I(x, y, t). Then for the pixel at (x, y) at time t, it moves to (x+dx, y+dy) at time t+dt, we have:
[0096] I(x+dx,y+dy,t+dt)=I(x,y,t) (4)
[0097] Based on the grayscale invariance assumption of the LK optical flow method, the grayscale at the next moment is equal to the grayscale at the previous moment, so equation (4) is further transformed into:
[0098]
[0099] In the above formula, is the gradient I of the image on the x-axis at that point x ; is the gradient I of the image on the y axis at that point y ; is the speed v of the pixel on the x-axis x ; is the velocity v on the y-axis y ;
[0100] The change of image grayscale with respect to time is recorded as I t ,get:
[0101]
[0102] Transform the above formula (6) into x , v y The overdetermined linear equation is solved using the least squares method to obtain the x and y velocities, thereby determining the dynamics at (x, y);
[0103] Step S42: dynamic attribute determination based on RANSAC;
[0104] Match the feature points of the current frame with the feature points of the reference frame to establish a correspondence between the two frames; use the RANSAC algorithm to estimate the motion model between the two frames to obtain the motion posture of the feature points;
[0105] In the process of motion estimation, RANSAC determines which matching points are inliers, i.e. points that conform to the motion model, and which are outliers, i.e. points that do not conform to the motion model, based on the error between the motion model and the matching points.
[0106] Step S43: Based on the method in step S41 and step S42, dynamic attribute determination is performed on the obtained active target and the passive target area having a subordinate relationship with the active target, and feature points in the area where they are located and are not static target areas are eliminated.
[0107] Specifically, step S5 specifically includes the following steps:
[0108] Step S51, combining YOLO, LK optical flow method, and RANSAC to determine the dynamic properties of active targets and passive target areas having a subordinate relationship therewith;
[0109] Step S52, removing feature points in the active target and passive target areas that are finally determined to be dynamic and not in the static target area;
[0110] Step S53: After removing the dynamic feature points, the static feature points are used to complete the subsequent operations of the ORB-SLAM2 algorithm.
[0111] The method of the present invention is further illustrated by a simulation experiment below.
[0112] Simulation environment: GPUNV IDIARTX3060, CPU i7-12700H, Ubuntu 20.04, CUDA 11.0, Pytorch 1.8.1.
[0113] The dataset is evaluated using the fr3_walking_xyz sequence in the public dataset Tum Dynamic Objects, and the entire process is dynamic.
[0114] Figure 2Absolute TrajectoryError, ATE, is the absolute trajectory error map in the fr3_walking_xyz data set of the present invention; Figure 3 This is the relative trajectory error map Relative Trajactory Error, RTE, in the fr3_walking_xyz dataset of the present invention.
[0115] The recognition accuracy of the method of the present invention is compared with that of the traditional ORB-SLAM2 and DS-SLAM methods in the fr3_walking_xyz data set, and the results are shown in Tables 1 to 4 below.
[0116] Table 1. Absolute trajectory errors of ORB-SLAM2 and the method of the present invention
[0117]
[0118] Table 2. Absolute trajectory error RMSE of DS-SLAM2 and the method of the present invention
[0119] sequence name DS-SLAM2 Method of the present invention Improved accuracy fr3_walking_xyz 0.0247 0.0151 38.9%
[0120] Table 3. Relative trajectory error translation part of ORB-SLAM2 and the method of the present invention
[0121]
[0122] Table 4. Relative trajectory error rotation part of ORB-SLAM2 and the method of the present invention
[0123]
[0124] It can be seen from the comparison results in the above table that the method of the present invention has a significant effect of improving accuracy compared with the traditional ORB-SLAM2 and DS-SLAM methods.
[0125] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.
Claims
1. A visual SLAM method based on dynamic attribute cognition, characterized in that: The following steps are involved: Step S1, perform ORB (feature point extraction) on the image to obtain the feature points of the image; Step S2: Use YOLO to perform target recognition on the image, detect various targets in the image, and input the two-dimensional position information of various targets into the ORB-SLAM2 algorithm; Step S3, performing affiliation determination on the passive targets around the active target to obtain the passive targets related to the active target; Step S4, performing dynamic attribute determination using the LK optical flow method and random sampling consistency RANSAC on the obtained active target and the passive target region having a subordinate relationship with the active target; Step S5: Remove the feature points in the active target and passive target areas that are finally determined to be dynamic and not in the static target area, and perform subsequent operations of ORB-SLAM2 on the remaining static feature points.
2. A visual SLAM method based on dynamic attribute cognition according to claim 1, characterized in that, The step S1 of extracting ORB feature points from an image comprises the following steps: Step S11, extracting feature points from the current image, and when the number of detected feature points exceeds a set threshold, initialization is considered successful; Step S12: Calculate the main direction of the ORB feature point using the grayscale centroid method, and define the moment m of the image block as: m pq =Σ x,y∈B x p y q I(x,y),p,q={0,1} (1) In the above formula: p is the offset of the pixel; q is the offset of the pixel; x is the horizontal coordinate of the pixel; y is the vertical coordinate of the pixel; B is the image block area; I (x, y) is the gray value of this coordinate; Find the centroid C of the image block through the moment m: Connect the geometric center O and the centroid C of the image block to obtain a direction vector OC, which defines the direction of the feature point as: Then, ORB feature points with direction information are obtained.
3. A visual SLAM method based on dynamic attribute cognition according to claim 1, characterized in that, In step S2, YOLO is used to identify the target in the image, detect various targets in the image, and input the two-dimensional position information of various targets into the ORB-SLAM2 algorithm. The process is as follows: Step S21, train YOLO using the COCO2017 dataset to obtain training weights, and use the trained YOLO to perform target recognition on the image; Step S22: Classify the target according to the target recognition result into the following three types: a. Static target: an object that does not move for a long time and is difficult to move; b. Active target: an entity that can move autonomously; c. Passive target: an object that moves due to the action of the active target; Based on the three types, dynamic attributes are divided into each target. Static objects are considered static unless there are special circumstances. Active targets are divided into transient and dynamic, and passive targets are divided into static, transient and dynamic. Among them, static refers to a state of not moving for a period of time; transient refers to a state with a probability of movement; dynamic refers to a state in motion; Step S23: input the two-dimensional position information of the target after classification in step S22 into the ORB-SLAM2 algorithm.
4. A visual SLAM method based on dynamic attribute cognition according to claim 1, characterized in that, In step S3, the passive targets around the active target are judged to have a subordinate relationship, and the passive targets related to the active target are obtained. The process is as follows: Step S31, according to the various targets in the image detected in step S2, the active targets in the image are processed in sequence, firstly, the targets that are not human in the active targets are determined, and the Euclidean distance between the center point of the active target and the center points of the surrounding passive targets is directly used to determine whether the distance is less than a set threshold value, and when the Euclidean distance is less than the set threshold value, it is determined that there is a subordinate relationship between the two; Step S32: When the active target is a human, additionally identify the key points of the human hand, and determine the distance between the passive target in the image and the key points of the hand. When the distance is less than the set threshold, it is determined that the passive target and the active target have a subordinate relationship and have the same dynamic properties.
5. A visual SLAM method based on dynamic attribute cognition according to claim 1, characterized in that: The dynamic attribute determination process in step S4 is as follows: Step S41, dynamic attribute determination based on LK optical flow method; Let the grayscale of a pixel at (x, y) at time t be recorded as I(x, y, t). Then for the pixel at (x, y) at time t, it moves to (x+dx, y+dy) at time t+dt, we have: I(x+dx,y+dy,t+dt)=I(x,y,t) (4) Based on the grayscale invariance assumption of the LK optical flow method, the grayscale at the next moment is equal to the grayscale at the previous moment, so equation (4) is further transformed into: In the above formula, is the gradient I of the image on the x-axis at that point x ; is the gradient I of the image on the y axis at that point y ; is the speed v of the pixel on the x-axis x ; is the velocity v on the y-axis y ; The change of image grayscale with respect to time is recorded as I t ,get: Transform the above formula (6) into x , v y The overdetermined linear equation is solved using the least squares method to obtain the x and y velocities, thereby determining the dynamics at (x, y); Step S42: dynamic attribute determination based on RANSAC; Match the feature points of the current frame with the feature points of the reference frame to establish a correspondence between the two frames; use the RANSAC algorithm to estimate the motion model between the two frames to obtain the motion posture of the feature points; In the process of motion estimation, RANSAC determines which matching points are inliers, i.e. points that conform to the motion model, and which are outliers, i.e. points that do not conform to the motion model, based on the error between the motion model and the matching points. Step S43: Based on the methods in step S41 and step S42, dynamic attributes of the obtained active target and the passive target area having a subordinate relationship with the active target are determined, and feature points in the area where they are located and are not static target areas are eliminated.
6. A visual SLAM method based on dynamic attribute cognition according to claim 1, characterized in that: Step S5 specifically includes the following steps: Step S51, combining YOLO, LK optical flow method, and RANSAC to determine the dynamic properties of active targets and passive target areas having a subordinate relationship therewith; Step S52, removing feature points in the active target and passive target areas that are finally determined to be dynamic and not in the static target area; Step S53: After removing the dynamic feature points, the static feature points are used to complete the subsequent operations of the ORB-SLAM2 algorithm.
Citation Information
Patent Citations
Image dynamic feature removing method and device
CN112037261A
Dynamic vision SLAM (Simultaneous Localization and Mapping) method based on target detection and feature point speed constraint
CN115115672A
Accurate positioning method for wheeled mobile robot
CN115540851A
SLAM dynamic point elimination method based on YOLO
CN116664631A
Anti-jitter real-time target tracking method based on LK optical flow method
CN117440246A
