Visual SLAM method for feature point processing in dynamic environment

Abstract

The invention discloses a visual SLAM method for feature point processing in a dynamic environment, and belongs to the field of visual SLAM, and the method comprises the steps: S1, constructing a feature point mismatching processing module and a dynamic feature point processing module for a feature point module in a visual SLAM system; s2, establishing a fusion algorithm for removing mismatched feature points in two stages in a feature point mismatching processing module, and removing the mismatched feature points by setting a dynamic threshold value; and S3, constructing a dynamic feature point distinguishing and removing method in the dynamic feature point processing module, segmenting a suspected dynamic region through a physical prior method, and judging and removing points in the region through an optical flow method. According to the method, the calculation amount for judging the points is reduced, some misjudgment points caused by external factor interference can be reserved as much as possible, and the performance of the system is improved; simulation experiment results show that the method has the outstanding advantages of being good in removal effect, excellent in physical performance and the like.

Description

Technical Field

[0001] The invention relates to the field of visual SLAM, in particular to a visual SLAM method for feature point processing in a dynamic environment. Background Art

[0002] In recent years, autonomous underwater vehicles (AUVs) have become increasingly important for ocean research and exploration. Accurate positioning and navigation are key to ensuring that AUVs can successfully complete their missions. Underwater positioning and mapping (SLAM) can build maps using environmental information captured by visual sensors in unknown underwater environments to locate AUVs. However, in complex underwater scenes with weak visual image textures and low light levels, SLAM systems have difficulty handling the problem of dynamic object intrusion that occurs during feature matching, making it a challenge to implement accurate underwater positioning and navigation.

[0003] Dynamic factors in the underwater environment, such as ocean currents, fish schools, and floating objects, can block static feature points in the SLAM system, causing underwater robot positioning and tracking failure. With the rapid development of deep learning technology, many visual SLAM technologies based on it have emerged. Although these methods enhance the system's stability to environmental changes and significantly improve the system's accuracy, these methods also have limitations:

[0004] 1. In the process of implementing semantic segmentation, inaccurate segmentation may occur, resulting in incomplete exclusion of dynamic points, which reduces the overall accuracy of positioning and navigation using the SLAM system. Therefore, in underwater dynamic environments, how to process dynamic feature points is the key to improving the positioning accuracy and robustness of visual SLAM.

[0005] 2. The existence of mismatched points in the system will seriously affect the accuracy and robustness of the visual SLAM system, resulting in a significant reduction in the positioning and navigation effect. Although there are some methods to deal with the mismatch problem between feature points, these methods only deal with the mismatch problem and do not consider the impact of dynamic objects in the system.

[0006] Therefore, in order to maximize the performance of the SLAM system used for AUV positioning and navigation, we need to comprehensively consider and solve the interference of dynamic factors in the scene and the mismatching problem between feature points. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a visual SLAM method for feature point processing in a dynamic environment, which can effectively solve the mismatch problem between feature points and reduce the influence of dynamic feature points on the overall accuracy of the system, thereby improving the positioning and navigation accuracy of AUVs.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] A visual SLAM method for feature point processing in a dynamic environment comprises the following steps:

[0010] S1. Construct a feature point mismatch processing module and a dynamic feature point processing module for the feature point module in the visual SLAM system;

[0011] S2, establishing a two-stage fusion algorithm for removing mismatched feature points in the feature point mismatch processing module, and removing mismatched points by setting a dynamic threshold;

[0012] S3. Construct a dynamic feature point identification and removal method in the dynamic feature point processing module, segment the suspected dynamic area through the physical prior method, and use the optical flow method to judge and remove the points in the area.

[0013] The further improvement of the technical solution of the present invention is that: in S1, in the ORB-SLAM3 system belonging to the visual SLAM system, for the feature point module, a feature point mismatch processing module and a dynamic feature point processing module are respectively constructed to form a system framework;

[0014] The system framework includes a feature point mismatch processing module and a dynamic feature point processing module; after extracting ORB feature points from the input adjacent frame images, the feature points in the reference frame and the current frame are initially matched; the feature point mismatch module will screen the preliminary matching results and remove unmatched feature point pairs to reduce unmatched feature point pairs; the dynamic feature point module processes dynamic objects appearing in the image; by comparing the feature points of the previous and next frames, the feature points belonging to the dynamic category area are eliminated, and then the remaining static feature points are used for feature point matching; the subsequent processing is consistent with the ORB-SLAM3 system; after entering the key frame judgment module, the system will enter the subsequent local mapping process and the loop and mapping merge thread.

[0015] A further improvement of the technical solution of the present invention is that S2 specifically includes the following steps:

[0016] After the S21 system extracts the ORB feature points of two adjacent frames of images, it first uses the BFMatch algorithm to obtain two sets M containing all the feature points in the two frames of images. 1 and M 2 ; After obtaining the original point set M 1and M 2 After that, all matching results are preliminarily screened using the minimum Hamming distance; a dynamic threshold k is set to improve its accuracy; the selection of the minimum Hamming distance is as follows:

[0017] k = min(2×min_dist, t)

[0018] HM_matches = {m i ∈M 1 ,m j ∈M 2 |d(m i ,m j )≤k}

[0019] Among them, t is the manually set threshold, min_dist is the minimum distance between feature points matching adjacent images, that is, the distance between correctly matched point pairs, and m i , m j Respectively represent M 1 and M 2 For the corresponding points in , HM_matches indicates that the feature point pairs are screened by the minimum Hamming distance;

[0020] S22 is screened by Hamming distance to form a new feature point N 1 and N 2 ,The RANSAC algorithm is used to further filter the feature points and remove the unmatched feature points;

[0021] The RANSAC algorithm screening process is as follows:

[0022] RANSAC_matches = {p i ∈N 1 ,q j ∈N 2 |||T(p i )-q j ||≤d}

[0023] Among them, p i ,q j Respectively represent N 1 and N 2 The points that match each other in 1 The pi points in the dataset are mapped to N 2 Coordinate transformation in the dataset, d represents the manually set threshold; RANSAC_matches represents feature point pairs screened by the RANSAC algorithm;

[0024] S23 is further screened by the RANSAC algorithm to form a point set I that matches the feature points in the adjacent frames more accurately. 1 and I 2, which is expressed as follows:

[0025] f(I 1 ,I 2 )={(n i ,n j )|n i ∈I 1 ,n j ∈I 2 ,HM_matches,RANSAC_matches}

[0026] Among them, n i and n j Respectively represent the point set I 1 and I 2 The feature points in 1 ,I 2 ) represents the points after removing the unmatched feature points.

[0027] A further improvement of the technical solution of the present invention is that S3 specifically includes the following steps:

[0028] S31 generates a corresponding motion vector when a feature point moves between adjacent frames;

[0029] Δr=r(t+Δt)-r(t)

[0030] Wherein, Δr represents the position change caused by the movement of the feature point, r(t) represents the position vector of the feature point at time t, Δt represents the change in time, and r(t+Δt) represents the position vector of the feature point after the time interval t+Δt.

[0031] By analyzing these motion vectors, these motion feature points can be divided;

[0032] S32 uses the motion prior segmentation function to segment the regions of these dynamic feature points, and then uses the optical flow method to further judge and eliminate them;

[0033] Assume that objects in an image sequence remain constant in brightness as they move between frames:

[0034] I(x,y,t)=I(x+u,y+v,t+Δt)

[0035] Among them, I(x,y,t) represents the brightness of the feature point at time t; Δt represents the change in time, I(x+u,y+v,t+Δt) represents the brightness of the same point after Δt; u and v represent the horizontal and vertical components of the optical flow;

[0036] S33 calculates the derivative of the image brightness function I with respect to time t:

[0037]

[0038] Under the assumption of constant brightness, the optical flow constraint equation is derived:

[0039]

[0040] In order to obtain the optical loss (u, v), the spatial gradient of the current frame needs to be calculated:

[0041]

[0042] Among them, ▽I represents the gradient of brightness I;

[0043] Convert it into the following matrix form:

[0044]

[0045] Among them, I x express I y express

[0046] S34 obtains the optical flow vector (u, v) of the segmented feature point and calculates the size S of the optical flow velocity vector of the feature point:

[0047]

[0048] Among them, u represents the optical flow component of the feature point in the horizontal direction, and v represents the optical flow component of the feature point in the vertical direction;

[0049] In order to remove dynamic feature points, a threshold S is set. max ; If the above S exceeds the threshold, it is judged as a dynamic feature point and is removed.

[0050] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0051] 1. The present invention designs a fusion algorithm for removing mismatched feature points, performs two-stage removal of mismatched feature points in the system, and introduces an adaptive dynamic threshold to prevent incomplete removal of mismatched points due to improper threshold setting, thereby achieving the technical effect of accurately and efficiently removing mismatched feature points in the system. Compared with a single removal algorithm, the method of the present invention has a significant improvement in the proportion of correct matching point pairs.

[0052] 2. The present invention combines physical priors and optical flow methods to segment and remove dynamic areas in the motion area, and at the same time retains those misjudged points caused by external interference through threshold discrimination, thereby achieving accurate and efficient removal of dynamic feature points, and retaining those misjudged points to the maximum extent, preventing the occurrence of too few feature points for positioning and mapping. Compared with the current mainstream SLAM algorithm, the present invention has a significant improvement in absolute trajectory error.

[0053] 3. In the process of removing mismatched points, the present invention introduces a dynamic threshold to avoid the problem of residual mismatched points caused by improper setting of fixed thresholds, thereby improving the accuracy and robustness of removing mismatched points.

[0054] 4. The present invention can more accurately identify and remove dynamic feature points through the combination of physical prior segmentation and optical flow method, reduce the interference of dynamic objects on the positioning and navigation system, and improve the environmental adaptability of the system.

[0055] 5. The present invention modularizes the processing of mismatched points and dynamic feature points, which facilitates the expansion and maintenance of the system and improves the operating efficiency and real-time performance of the system.

[0056] 6. The present invention has been verified through experiments and optimized the threshold setting for dynamic feature point removal, ensuring that the best dynamic feature point removal effect can be achieved in different environments, further improving the accuracy and stability of positioning navigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a system block diagram of the method for fusion processing of feature points in the present invention;

[0058] Figure 2 This is a schematic diagram of feature point mismatching in the present invention;

[0059] Figure 3 It is a schematic diagram of segmentation of dynamic feature points and misjudged dynamic feature points in the present invention. DETAILED DESCRIPTION

[0060] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:

[0061] like Figure 1 As shown, a visual SLAM method for feature point processing in a dynamic environment includes the following steps:

[0062] S1. Build a system framework for feature point processing based on visual SLAM system;

[0063] In the ORB-SLAM3 system belonging to the visual SLAM system, for the feature point module, a feature point mismatch processing module and a dynamic feature point processing module are constructed respectively to form a new framework to deal with these two problems. In this framework, the key frame image input into the system is firstly subjected to the fast feature detection and description algorithm (OrientedFAST and Rotated BRIEF, ORB) feature point extraction, and then the extracted feature points are processed by the two modules we proposed (feature point mismatch processing module and dynamic feature point processing module), and the processed feature points are output to participate in the tracking thread to generate new key frames, which are then involved in the subsequent local mapping thread and the loopback and map merging thread;

[0064] System framework such as Figure 1 As shown in the figure, it includes a feature point mismatch processing module and a dynamic feature point processing module; after extracting ORB feature points from the input adjacent frame images, the feature points in the reference frame and the current frame are initially matched. The feature point mismatch module will screen the preliminary matching results and remove unmatched feature point pairs to reduce unmatched feature point pairs. The dynamic feature point module processes dynamic objects appearing in the image. By comparing the feature points of the previous and next frames, the feature points belonging to the dynamic category area are eliminated, and then the remaining static feature points are used for feature point matching. The subsequent processing is consistent with the ORB-SLAM3 system. After entering the key frame judgment module, the system will enter the subsequent local mapping process and the loop and mapping merge thread.

[0065] S2, establishing a two-stage fusion algorithm for removing mismatched feature points in the feature point mismatch processing module;

[0066] Feature matching, as a key technology in computer vision, has been widely used in many fields. Feature matching is the process of extracting and comparing local features from two images obtained from different angles of the same scene to determine the correspondence between the images. In traditional algorithms for removing mismatched feature points, these mismatched points are only removed once, resulting in incomplete removal of these points, which leads to the existence of corresponding errors in the system and affects the system accuracy. The Scale Invariant Feature Transform (SIFT) and Speeded-Up Robust Features (SURF) features are two high-dimensional feature descriptors in the feature matching process, which have very high requirements for computing power. However, the onboard hardware and software resources of mobile robots limit this capability. ORB is a fast feature point extraction and description algorithm. By combining the Features from Accelerated Segment Test (FAST) feature point detection method with the Binary Robust Independent Elementary Features (BRIEF) feature descriptor, it can provide more complete feature matching efficiency. However, the current ORB feature matching algorithm produces many false feature matches in some challenging scenes, such as lighting changes, motion blur, image rotation, viewpoint changes, and weak texture scenes. The purpose of this module is to solve the problem of inaccurate feature matching in visual navigation systems.

[0067] Therefore, in order to achieve accurate and large-scale removal of these mismatched points, a fusion algorithm combining BFMatch and RANSAC is proposed, aiming to reduce the mismatch rate and increase the correct matching point pairs;

[0068] S2 specifically includes the following steps:

[0069] After the S21 system extracts the ORB feature points of two adjacent frames of images, there are mismatching problems between feature points with small pixel differences, such as Figure 2 As shown in the figure, the correct matching point corresponding to point a is a′, but in actual situations, point a may match point b′, point c′, point d′, point e′, and point f′. To solve this problem, we first use the BFMatch algorithm to obtain two sets M containing all the feature points in the two frames. 1 and M 2 ; After obtaining the original point set M 1 and M 2Finally, all matching results are preliminarily screened using the minimum Hamming distance. In order to avoid the problem of low screening accuracy in this step, a dynamic threshold k is set to improve its accuracy. The selection of the minimum Hamming distance is as follows:

[0070] k = min(2×min_dist, t)

[0071] HM_matches = {m i ∈M 1 ,m j ∈M 2 |d(m i ,m j )≤k}

[0072] Among them, t is the manually set threshold, min_dist is the minimum distance between feature points matching adjacent images, that is, the distance between correctly matched point pairs, and m i , m j Respectively represent M 1 and M 2 In the figure, HM_matches indicates that the feature point pairs are screened by the minimum Hamming distance. In order to prevent the incomplete screening of mismatched feature points due to improper setting of the threshold t, a dynamic threshold k is introduced to make k reasonably change between different pairs, thus reducing the problem that mismatched points are considered as correct matching points due to improper threshold setting.

[0073] S22 is screened by Hamming distance to form a new feature point N 1 and N 2 ,The RANSAC algorithm is used to further filter the feature points and remove the unmatched feature points;

[0074] The RANSAC algorithm screening process is as follows:

[0075] RANSAC_matches = {p i ∈N 1 ,q j ∈N 2 |||T(p i )-q j ||≤d}

[0076] Among them, p i ,q j Respectively represent N 1 and N 2 The points that match each other in 1 The pi points in the dataset are mapped to N 2 Coordinate transformation in the dataset, d represents the manually set threshold; RANSAC_matches represents feature point pairs screened by the RANSAC algorithm;

[0077] S23 is further screened by the RANSAC algorithm to form a point set I that matches the feature points in the adjacent frames more accurately. 1 and I 2 , which is expressed as follows:

[0078] f(I 1 ,I 2 )={(n i ,n j )|n i ∈I 1 ,n j ∈I 2 ,HM_matches,RANSAC_matches}

[0079] Where n i and n j Respectively represent the point set I 1 and I 2 The feature points in 1 ,I 2 ) represents the points after removing the unmatched feature points.

[0080] S3, constructing a dynamic feature point identification and removal method in a dynamic feature point processing module;

[0081] The ORB-SLAM3 algorithm performs positioning, navigation and map construction based on the assumption of a static environment, without considering the interference of dynamic targets. However, in the real marine environment, there are many dynamic objects, such as marine life, floating marine debris and underwater detection equipment, which makes the assumption of a static environment invalid, making the final positioning and navigation effects unsatisfactory. The application of physical priors in computer vision mainly includes image segmentation, target tracking, 3D reconstruction and anomaly detection. In image segmentation, motion priors can provide clues about the changes of objects over time. For example, in a video sequence, the movement of an object is usually more consistent and predictable than the background. Using this method, segmentation can be achieved by analyzing the motion vectors of feature points between consecutive frames. Optical flow is a method for estimating the motion of pixels between two consecutive images. In the real world, the movement of an object causes the position of the corresponding pixel in the image to change. From these changes, the speed and direction of the object can be estimated. The optical flow method is used to obtain the speed change of the feature points of the moving target in adjacent frames. When it exceeds a certain threshold, it will be eliminated. Therefore, this module proposes a method that combines physical priors and optical flow to segment and eliminate the feature points of the identified dynamic objects.

[0082] Dynamic feature points will cause serious interference to the positioning and navigation performance of underwater robots. If the underwater robot using visual SLAM for positioning and navigation does not consider the influence of dynamic feature points, the positioning and navigation performance of the underwater robot will be very inaccurate.

[0083] In order to solve this problem, a method combining physical prior segmentation with optical flow is proposed to segment and identify the dynamic feature point area, thereby effectively reducing the impact of dynamic feature points on the system to achieve efficient and accurate positioning and navigation, including the following steps:

[0084] S31 generates a corresponding motion vector when a feature point moves between adjacent frames;

[0085] Δr=r(t+Δt)-r(t)

[0086] Wherein, Δr represents the position change caused by the movement of the feature point, r(t) represents the position vector of the feature point at time t, Δt represents the change in time, and r(t+Δt) represents the position vector of the feature point after the time interval t+Δt.

[0087] By analyzing these motion vectors, these motion feature points can be divided;

[0088] S32 uses the motion prior segmentation function to segment the regions of these dynamic feature points, and then uses the optical flow method to further judge and eliminate them;

[0089] When using the optical flow method to eliminate feature points, the premise of constant brightness should be met; this means that if an object moves between two frames, the brightness of its pixels at the new position is the same as the brightness of its pixels at the original position; therefore, we assume that the brightness of objects in the image sequence remains unchanged when they move between different frames:

[0090] I(x,y,t)=I(x+u,y+v,t+Δt)

[0091] Among them, I(x,y,t) represents the brightness of the feature point at time t; Δt represents the change in time, I(x+u,y+v,t+Δt) represents the brightness of the same point after Δt; u and v represent the horizontal and vertical components of the optical flow;

[0092] S33 calculates the derivative of the image brightness function I with respect to time t:

[0093]

[0094] Under the assumption of constant brightness, the optical flow constraint equation is derived:

[0095]

[0096] In order to obtain the optical loss (u, v), the spatial gradient of the current frame needs to be calculated:

[0097]

[0098] Among them, ▽I represents the gradient of brightness I;

[0099] Convert it into the following matrix form:

[0100]

[0101] Among them, I x express I y express

[0102] S34 obtains the optical flow vector (u, v) of the segmented feature point and calculates the size S of the optical flow velocity vector of the feature point:

[0103]

[0104] Among them, u represents the optical flow component of the feature point in the horizontal direction, and v represents the optical flow component of the feature point in the vertical direction;

[0105] In order to remove dynamic feature points, we set a threshold S max ; If the above S exceeds the threshold, it is judged as a dynamic feature point and is removed.

[0106] Through the above method, we segment the areas that may be dynamic feature points for processing, such as Figure 3 As shown in the figure, the red area is the dynamic feature point area, and the green area is the suspected dynamic point area. We divide them out using the physical prior method, avoiding the subsequent optical flow method to solve the optical flow vectors of all feature points, which greatly saves the calculation time. At the same time, by setting a reasonable threshold S max , it can avoid non-dynamic feature points caused by external factors being misjudged as dynamic feature points, eliminate this situation, retain static feature points to the greatest extent, reduce the impact of external factors, and thus effectively reduce the impact on positioning and navigation accuracy.

[0107] Example 1: Experimental verification of feature point mismatching in a water pool

[0108] In order to experimentally verify the feature point mismatching module we proposed, we conducted relevant experimental verification in a pool simulating an ocean environment. The verification results are shown in Table 1:

[0109] Table 1

[0110]

[0111]

[0112] In Table 1, we can clearly see that when the total number of feature points is the same, our proposed method is the best one in dealing with the mismatch problem between feature points compared with other methods.

[0113] Example 2: Experimental verification of dynamic feature point processing

[0114] In the experimental verification of dynamic feature point processing, we used the München, TUM) dataset, and compared with several current mainstream dynamic feature point removal methods to verify the performance advantages of our proposed method. The absolute trajectory error is shown in Table 2, and the relative pose error is shown in Table 3:

[0115] Table 2 ATE

[0116]

[0117] Table 3 RPE

[0118]

[0119] It can be seen from Tables 2 and 3 that on the TUM dataset, the method we proposed is superior to the current mainstream dynamic feature point removal method in terms of both absolute trajectory error and relative pose error, and can minimize the impact of dynamic feature points on the underwater robot when it uses visual SLAM for positioning and navigation.

[0120] The embodiments are merely the technical ideas of the invention described in the specification and cannot be used to limit the protection scope of the invention. Any changes made on the basis of the technical solutions in accordance with the technical ideas proposed by the invention shall fall within the protection scope of the invention.

Claims

1. A visual SLAM method for feature point processing in a dynamic environment, characterized by: The following steps are involved: S1. Construct a feature point mismatch processing module and a dynamic feature point processing module for the feature point module in the visual SLAM system; S2, establishing a two-stage fusion algorithm for removing mismatched feature points in the feature point mismatch processing module, and removing mismatched points by setting a dynamic threshold; S3. Construct a dynamic feature point identification and removal method in the dynamic feature point processing module, segment the suspected dynamic area through the physical prior method, and use the optical flow method to judge and remove the points in the area.

2. The visual SLAM method for feature point processing in a dynamic environment according to claim 1, characterized in that: In S1, in the ORB-SLAM3 system belonging to the visual SLAM system, for the feature point module, a feature point mismatch processing module and a dynamic feature point processing module are respectively constructed to form a system framework; The system framework includes a feature point mismatch processing module and a dynamic feature point processing module; after extracting ORB feature points from the input adjacent frame images, initially matching the feature points in the reference frame and the current frame; The feature point mismatch module will screen the preliminary matching results and remove the unmatched feature point pairs to reduce the number of unmatched feature point pairs. The dynamic feature point module processes dynamic objects in the image. By comparing the feature points of the previous and next frames, the feature points belonging to the dynamic category area are removed, and then the remaining static feature points are used for feature point matching. The subsequent processing is consistent with the ORB-SLAM3 system. After entering the key frame judgment module, the system will enter the subsequent local mapping process and the loop and mapping merging thread.

3. The visual SLAM method for feature point processing in a dynamic environment according to claim 1, characterized in that: S2 specifically includes the following steps: After the S21 system extracts the ORB feature points of two adjacent frames, it first uses the BFMatch algorithm to obtain two sets M1 and M2 containing all the feature points in the two frames. After obtaining the original point sets M1 and M2, it uses the minimum Hamming distance to preliminarily screen all matching results. It sets a dynamic threshold k to improve its accuracy. The selection of the minimum Hamming distance is as follows: k=min(2×min_dist,t) HM_matches={m i ∈M1,m j ∈M2|d(m i ,m j )≤k} Among them, t is the manually set threshold, min_dist is the minimum distance between feature points matching adjacent images, that is, the distance between correctly matched point pairs, and m i , m j They represent the corresponding points in M1 and M2 respectively, and HM_matches indicates that the feature point pairs are screened by the minimum Hamming distance; S22 is screened by Hamming distance to form new feature points N1 and N2, and the RANSAC algorithm is used to further filter the feature points to remove unmatched feature points; The RANSAC algorithm screening process is as follows: RANSAC_matches={p i ∈N1,q j ∈N2|||T(p i )-q j ||≤d} Among them, p i ,q j Respectively represent the matching points in N1 and N2, T represents the coordinate transformation of mapping the pi point in the N1 data set to the N2 data set, d represents the manually set threshold; RANSAC_matches means that the feature point pairs are screened by the RANSAC algorithm; S23 is further screened by the RANSAC algorithm to form point sets I1 and I2 that more accurately match the feature points in adjacent frames. The expressions are as follows: f(I1,I2)={(n i ,n j )|n i ∈I1,n j ∈I2,HM_matches,RANSAC_matches} Among them, n i and n j They represent the feature points in point sets I1 and I2 respectively, and f(I1,I2) represents the points after removing the non-matching feature points.

4. The visual SLAM method for feature point processing in a dynamic environment according to claim 1, characterized in that: S3 specifically includes the following steps: S31 generates a corresponding motion vector when a feature point moves between adjacent frames; Δr=r(t+Δt)-r(t) Wherein, Δr represents the position change caused by the movement of the feature point, r(t) represents the position vector of the feature point at time t, Δt represents the change in time, and r(t+Δt) represents the position vector of the feature point after the time interval t+Δt. By analyzing these motion vectors, these motion feature points can be divided; S32 uses the motion prior segmentation function to segment the regions of these dynamic feature points, and then uses the optical flow method to further judge and eliminate them; Assume that objects in an image sequence remain constant in brightness as they move between frames: I(x,y,t)=I(x+u,y+v,t+Δt) Among them, I(x,y,t) represents the brightness of the feature point at time t; Δt represents the change in time, I(x+u,y+v,t+Δt) represents the brightness of the same point after Δt; u and v represent the horizontal and vertical components of the optical flow; S33 calculates the derivative of the image brightness function I with respect to time t: Under the assumption of constant brightness, the optical flow constraint equation is derived: In order to obtain the optical loss (u, v), the spatial gradient of the current frame needs to be calculated: Among them, ▽I represents the gradient of brightness I; Convert it into the following matrix form: Among them, I x express I y express S34 obtains the optical flow vector (u, v) of the segmented feature point and calculates the size S of the optical flow velocity vector of the feature point: Among them, u represents the optical flow component of the feature point in the horizontal direction, and v represents the optical flow component of the feature point in the vertical direction; In order to remove dynamic feature points, a threshold S is set. max ; If the above S exceeds the threshold, it is judged as a dynamic feature point and is removed.

Images