A visual SLAM method based on multi-mask fusion

Through the multi-mask fusion visual SLAM method, semantic segmentation and geometric error are used to model feature points and generate static object maps, which solves the positioning and mapping accuracy problems of visual SLAM in dynamic environments and improves the robustness and accuracy of the system.

CN119295798BActive Publication Date: 2025-09-23ZHEJIANG UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411247201.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-06
Publication Date
2025-09-23
Estimated Expiration
2044-09-06

AI Technical Summary

Technical Problem

The existing visual SLAM system suffers from reduced positioning and mapping accuracy in dynamic environments, and problems such as tracking loss and relocalization failure occur, making it difficult to adapt to dynamic scenes.

Method used

A visual SLAM method based on multi-mask fusion is adopted. The RGB image is classified through a semantic segmentation network. The chi-square test is performed based on the epipolar constraint error and the depth projection error. The static weights of the feature points are calculated, the dynamic area mask is generated, the dynamic point cloud is eliminated, and a map containing only static objects is generated.

Benefits of technology

The robustness and accuracy of the visual SLAM system in dynamic scenes are improved, and the accurate positioning and mapping functions of the robot in dynamic environments are realized.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119295798B_ABST
    Figure CN119295798B_ABST
Patent Text Reader

Abstract

The present invention discloses a visual SLAM method based on multi-mask fusion, comprising: obtaining an RGB image and its corresponding depth image to extract feature points and depth information; classifying objects in the RGB image into C1, C2 and C3; performing a preliminary estimation of the camera pose using only the feature points in the C1 area to obtain T init , according to T init The epipolar constraint error and depth projection error of the feature points are calculated and fused together. The static feature points belonging to the C1 category are modeled, and the motion state of the remaining feature points is determined. The static weights of the feature points that are not dynamic points are calculated and used to weight the camera pose estimation to obtain the final camera pose estimation result. The dynamic area mask is calculated based on the dynamic points and fused with the dynamic area mask of the C3 category area to obtain the first dynamic area mask, which is then used to reconstruct the static scene. This method accurately identifies the moving areas and generates a map containing only static objects.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision, and in particular relates to a visual SLAM method based on multi-mask fusion. Background Art

[0002] In recent years, thanks to the continuous advancements in sensor and computer technology, robotics has been rapidly updated and widely applied, encompassing diverse fields such as the service industry, industry, and the military. Within the field of robotics, simultaneous localization and mapping (SLAM) technology plays a central role. However, most current visual SLAM research assumes a static environment, assuming the robot's operating environment is static and the lighting remains constant. In practical applications, the environment often contains dynamic objects, such as moving vehicles and pedestrians, which significantly impact the localization and mapping accuracy of visual SLAM systems. The presence of dynamic objects can lead to tracking loss, relocalization failures, and even affect the normal operation of the system. Feature points extracted by the system may be located on moving objects, and their positions change over time, causing errors in the system's pose estimation. Therefore, research on visual SLAM that can adapt to dynamic scenes is of great significance and practical value. Summary of the Invention

[0003] Dynamic objects can cause a significant decrease in the accuracy of visual SLAM positioning and mapping, resulting in problems such as tracking loss and relocalization failure. In order to achieve accurate positioning and mapping of robots in dynamic environments, the present invention proposes a visual SLAM method based on multi-mask fusion, which reduces the impact of dynamic objects in real scenes on the positioning and mapping of the visual SLAM method.

[0004] To achieve the above effects, the technical solution adopted by the present invention is as follows: a visual SLAM method based on multi-mask fusion, comprising the following steps:

[0005] S1. Obtain an RGB image and its corresponding depth image, extract feature points from the RGB image, and obtain depth information from the depth image;

[0006] S2. Classify the objects in the RGB image into C1, C2 and C3 respectively; C1 represents an absolutely stationary object, C2 represents a potentially moving object, and C3 represents a pre-set moving object;

[0007] S3, use only the feature points in the C1 area to perform a preliminary estimate of the camera pose and obtain T init , according to T init Calculate the epipolar constraint error e of the feature point epi and the depth projection error e depth, the two errors are fused, and the static feature points belonging to the C1 category are modeled, and then the chi-square test is used to determine the motion status of the remaining feature points;

[0008] S4. Calculate the static weight ω of the feature points that are not dynamic points i static , used to weight the camera pose estimation to obtain the final camera pose estimation result T final ;

[0009] S5. Calculate the dynamic area mask M based on the dynamic points g and the dynamic region mask M of the C3 category region s Fusion is performed to obtain the first dynamic area mask Based on the dynamic area detection result, the static scene is reconstructed.

[0010] Furthermore, the ORB method is used to extract feature points in the RGB image and match them with the feature points in the previous key frame for camera pose estimation.

[0011] Furthermore, the objects in the RGB image are classified into C1, C2 and C3, respectively, including:

[0012] The RGB image is preprocessed using a semantic segmentation network, and the objects in the image are classified into C1, C2, and C3 categories according to the labels of the semantic segmentation results.

[0013] Furthermore, the init Calculate the epipolar constraint error e of the feature point epi and the depth projection error e depth , the two errors are fused and the static feature points belonging to the C1 category are modeled. Then the chi-square test is used to determine the motion state of the remaining feature points, including:

[0014] Assume that the previous key frame and the current frame are and p1 and p2 are the matching points between the two frames, where p1 is the current frame A feature point in , p2 is the previous key frame A feature point in the , the homogeneous coordinates of these two points are p1 = [u1, v1, 1] T and p2 = [u2, v2, 1] T , then the epipolar constraint error of point p1 is:

[0015]

[0016] Among them, u1, v1, u2, v2 represent the coordinates of the feature point in the image coordinate system, X, Y are the vectors of the polar line of the current frame, F represents the basic matrix, and the superscript T represents the transpose; project the three-dimensional point P corresponding to the feature point p1 to the previous frame Get point p' and get its projected depth value z proj , point p′ in the frame The depth value in is z′; then the depth projection error of point p1 is:

[0017] e depth =z proj -z′

[0018] According to the preliminary estimated pose result T init , the average epipolar constraint error and depth projection error of the prior static points are calculated to be and And these two errors are independent and identically distributed, then the error distribution that the feature point in the static state should obey is:

[0019]

[0020] Among them, Σ is the covariance matrix, which represents the uncertainty of the scale. The calculation formula of the covariance matrix Σ of a feature point is:

[0021]

[0022] Among them, s represents the scale of each layer of the image pyramid extracted by the feature point, p represents the standard deviation of each layer of the image pyramid, and n represents the number of image pyramid layers where the feature point is located; the feature point polar constraint error e epi and the depth projection error e depth Combined into an error vector E = [e epi ,e depth ], for stationary feature points, the error vector composed of its errors should obey the Gaussian distribution, and its error vector E is expressed as:

[0023]

[0024] in, Represents the average error vector value of the feature points belonging to category C1. By calculating the inner product of the error vector E, the feature points p belonging to categories C2 and C3 are obtained. i Chi-square test value of :

[0025]

[0026] The obtained r i It should follow a chi-square distribution with 2 degrees of freedom:

[0027]

[0028] According to the feature point p i The chi-square test value r i Determine the motion state of the point:

[0029]

[0030] Furthermore, the calculation of the static weights of the feature points that are not dynamic points Used to weight the camera pose estimation to obtain the final camera pose estimation result T final ,include:

[0031] For feature points p that are not dynamic points i , its static weight The calculation formula is:

[0032]

[0033] in Indicates the feature point p that is not a dynamic point i The static weight value, r i Indicates the feature point p that is not a dynamic point i Chi-square test value, Semantic(p i ) represents a feature point p that is not a dynamic point i The label category is added to the static weight to obtain the final camera pose estimation:

[0034]

[0035] Where R and t represent the rotation matrix and translation vector of the camera, respectively, m represents the number of feature points that are not dynamic points, and Π(·) represents the camera projection model.

[0036] Furthermore, the dynamic area mask M is obtained by calculating the dynamic points. g ,include:

[0037] Only the dynamic points in the C2 category area are calculated to obtain the dynamic area mask M g .

[0038] Furthermore, the dynamic area mask M is obtained by calculating the dynamic points. g and the dynamic region mask M of the C3 category region s Fusion is performed to obtain the first dynamic area mask include:

[0039] According to the Euclidean distance between dynamic points, the DBSCAN clustering method is used to group them and obtain the convex hull of each group to generate the motion area mask M. gand the dynamic region mask M of the C3 category region s Fusion is performed to obtain the first dynamic area mask

[0040]

[0041] Furthermore, the step S5 further includes:

[0042] First dynamic area mask Divided into T independent masks And calculate each independent mask The average depth value d t :

[0043] According to d t Find the image area whose depth value is within the first preset threshold range

[0044] Will Decompose into K independent regions and calculate and The intersection rate between the two masks is determined by the intersection rate being greater than or equal to the second preset threshold. The effective mask in and will and Fusion obtains the second dynamic area mask H represents the number of valid masks.

[0045] Furthermore, the first dynamic area mask Divided into T independent masks Expressed as:

[0046]

[0047] The calculation where each independent mask The average depth value d t , the calculation formula is:

[0048]

[0049] Among them, p j Indicates The jth pixel in the region, D(p j ) represents point p j The depth value of

[0050] According to d t Find the image area whose depth value is within the first preset threshold range The judgment process is:

[0051]

[0052] Wherein, p represents any pixel in the image, and τ1 represents the first preset threshold;

[0053] Will Decompose into K independent regions Expressed as:

[0054]

[0055] and calculate and The intersection rate between two masks is:

[0056]

[0057] Where P(·,·) represents the intersection rate between the two masks, and N(·,·) represents the number of pixels that intersect between the two masks;

[0058] Determine by the intersection rate being greater than or equal to the second preset threshold τ2 The effective mask in q represents There are q conditions that meet the conditions area, represents one of the regions that meets the condition, 1≤z≤q;

[0059] The said and Fusion obtains the second dynamic area mask Expressed as:

[0060]

[0061] Furthermore, the reconstructing of the static scene based on the dynamic area detection result includes:

[0062] When building a dense point cloud map, the point cloud in the dynamic area is removed to obtain a static and reliable point cloud map.

[0063] Compared to existing methods, the present invention's main benefits lie in accurately identifying moving areas and generating maps containing only static objects. These innovations significantly improve the system's robustness in dynamic scenarios, enabling robots to accurately locate and map objects in these environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of a process of a specific embodiment of the present invention;

[0065] Figure 2Schematic diagram of the multi-mask fusion process in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0066] To make the purpose, design ideas, and technical solutions of the embodiments of the present invention clearer, the present invention is further described below with reference to the accompanying drawings. The features of the following embodiments and implementations can be combined with each other unless they conflict.

[0067] Reference Figure 1 As shown, the present invention proposes a visual SLAM method based on multi-mask fusion, which includes the following steps:

[0068] Step S1: Obtain an RGB image and its corresponding depth image, extract feature points from the RGB image, and obtain depth information from the depth image;

[0069] Specifically, an RGB-D camera acquires RGB and depth images of the current environment and extracts feature points from the RGB images. The ORB method (Oriented Fast and Rotated Brief) is used to extract feature points from the RGB image and match them with feature points from the previous keyframe for subsequent camera pose estimation. Keyframes are representative frames selected along the camera's trajectory. These frames store important information about the environment, such as feature points and pose, for subsequent map construction and localization.

[0070] Step S2: Classify the objects in the RGB image into C1, C2, and C3; C1 represents an absolutely stationary object, C2 represents a potentially moving object, and C3 represents a pre-set moving object;

[0071] Specifically, a semantic segmentation network is used to preprocess the RGB image. The objects in the image are then classified into C1, C2, and C3 categories based on the semantic segmentation results. In the object classification based on labels, C1 represents permanently stationary objects, such as buildings and the sky; C2 represents potentially moving objects, such as vehicles, chairs, and balloons; and C3 represents pre-defined moving objects, such as people and animals.

[0072] Step S3: Use only the feature points in the C1 region to perform a preliminary estimation of the camera pose and obtain T init , according to T init Calculate the epipolar constraint error e of the feature point epi and the depth projection error e depth , the two errors are fused, and the static feature points belonging to the C1 category are modeled, and then the chi-square test is used to determine the motion status of the remaining feature points.

[0073] Specifically, we first use the feature points belonging to the C1 category to perform initial pose estimation and obtain T init . And according to T init Calculate the epipolar constraint error e of the feature point epi and depth projection error e depth . Assume that the previous key frame and the current frame are and p1 and p2 are the matching points between the two frames, where p1 is the current frame A feature point in the frame, p2 is the previous frame A feature point in the , the homogeneous coordinates of these two points are p1 = [u1, v1, 1] T and p2 = [u2, v2, 1] T , where u1, v1, u2, v2 represent the coordinates of the feature point in the image coordinate system. The epipolar constraint error of point p1 is:

[0074]

[0075] Among them, X, Y are the vectors of the polar lines of the current frame, F represents the basic matrix, and the superscript T represents the transpose. Project the three-dimensional point P corresponding to the feature point p1 to the previous frame Get point p' and get its projected depth value z proj , point p′ in the frame The depth value in is z′. Then the depth projection error of point p1 is:

[0076] e depth =z proj -z′(2)

[0077] According to the semantic classification results, a set of prior static feature points (feature points in the C1 category) is obtained, and based on the preliminary estimated pose result T init , the average epipolar constraint error and depth projection error of the prior static point can be calculated as and Assuming that the two errors are independent and identically distributed, the error distribution that the feature points in a stationary state should obey is:

[0078]

[0079] Among them, Σ is the covariance matrix, which represents the uncertainty of the scale. The calculation formula of the covariance matrix Σ of a feature point is:

[0080]

[0081] Among them, s represents the scale of each layer of the image pyramid extracted by the feature point, p represents the standard deviation of each layer of the image pyramid, and n represents the number of image pyramid layers where the feature point is located. epi and the depth projection error e depth Combined into an error vector E = [e epi ,e depth ], for stationary feature points, the error vector composed of its errors should obey the Gaussian distribution, and its error vector E can be expressed as:

[0082]

[0083] in, Represents the average error vector value of the feature points belonging to category C1. By calculating the inner product of the error vector E, the feature points p belonging to categories C2 and C3 are obtained. i Chi-square test value of :

[0084]

[0085] The obtained r i It should follow a chi-square distribution with 2 degrees of freedom:

[0086]

[0087] According to the feature point p i The chi-square test value r i Determine the motion state of the point:

[0088]

[0089] The motion state of the feature points will be used for camera pose estimation in the next step.

[0090] Step S4: Calculate the static weights of feature points that are not dynamic points Used to weight the camera pose estimation to obtain the final camera pose estimation result T final ;

[0091] In particular, semantics and geometric constraints are combined to calculate the weights of feature points. For feature points p that are not dynamic points, i , its static weight The calculation formula is:

[0092]

[0093] in Indicates that it does not belong to the dynamic point p i The static weight value, r i Indicates that it does not belong to the dynamic point p i Chi-square test value, Semantic(pi ) indicates that it does not belong to the dynamic point p i The label category obtained in semantic segmentation. Add the obtained static weight to the final camera pose estimation to obtain:

[0094]

[0095] Where R, t represent the rotation matrix and translation vector of the camera respectively, m represents the number of feature points that are not dynamic points, Π(·) represents the camera projection model, and i represents the i-th map point

[0096] Step S5: Calculate the dynamic area mask M based on the dynamic points g and the dynamic region mask M of the C3 category region s Fusion is performed to obtain the first dynamic area mask Based on the dynamic area detection result, the static scene is reconstructed.

[0097] Preferably, only the dynamic points in the C2 category area are calculated to obtain the dynamic area mask M g .

[0098] In one embodiment, if Figure 2 As shown, the dynamic area mask M is calculated based on the dynamic points g and the dynamic region mask M of the C3 category region s Fusion is performed to obtain the first dynamic area mask include:

[0099] According to the Euclidean distance between dynamic points, the DBSCAN clustering method is used to group them and obtain the convex hull of each group, and then generate the motion area mask M obtained by geometric information g , and compare it with the mask M of the C3 category area obtained by semantic segmentation s Fusion is performed to obtain the first dynamic area mask for:

[0100]

[0101] In another embodiment, Figure 2 As shown, it includes two stages. In the first stage, the first dynamic area mask is obtained from the above embodiment. Phase II includes:

[0102] According to the connectivity between pixels, the first dynamic area mask is divided into T independent masks to facilitate the calculation of the average depth value of each different motion area.

[0103]

[0104] in, express The tth independent mask in , T means there are T masks in total. First calculate the average depth value of each independent mask in the depth image, assuming that There are J pixels in The average depth value d of the area t for:

[0105]

[0106] Among them, d t express The average depth value of the region, p j Indicates The jth pixel in the region, D(p j ) represents point p j The depth value of d t Find the image area with a depth value close to the image area (i.e. within the first preset threshold range), and the judgment process is as follows:

[0107]

[0108] Wherein, p represents any pixel in the image; τ1 represents a first preset threshold, also called a depth threshold (in one embodiment, its value is 0.1 m, but this is not a limitation), Indicates that in the depth image There are regions with similar depth values. Assume It can be decomposed into K independent regions and the intersection rate between the two masks is calculated:

[0109]

[0110] in, express In the kth independent region, P(·,·) represents the intersection rate between the two masks, and N(·,·) represents the number of pixels intersecting between the two masks. A second preset threshold τ2 is set (in one embodiment, its value is 0.2, but it is not limited to this), and the depth mask obtained based on the intersection rate is judged. Assume that There are q conditions that meet the conditions The second dynamic area mask It can be expressed as:

[0111]

[0112]

[0113] in, Represents a valid mask, there are H valid masks in total; H≤T, T is the first dynamic area mask Separate masks Formula (19) represents the fusion of the mask obtained in the first stage and the mask obtained in the second stage to obtain the final mask result

[0114] The first dynamic area mask can be obtained from this and the second motion region mask Preferred second motion region mask When building a dense point cloud map, the point cloud in the moving area can be removed to obtain a static and reliable point cloud map.

[0115] The present invention employs a multi-mask fusion-based visual SLAM method to improve the robustness and accuracy of positioning and mapping in dynamic scenes. This method utilizes semantic segmentation to extract a priori static information from the scene and models feature points in static regions by combining multiple geometric errors. A chi-square test is then employed to adaptively detect dynamic feature points in environments with varying degrees of dynamicity. To reduce the impact of system noise on pose estimation, a method for calculating static weights for feature points is designed, leading to a pose optimization method based on static weights. Furthermore, to accurately identify dynamic regions, this method fuses dynamic point detection results with semantic information to preliminarily identify moving objects. Dynamic region masks are then refined using depth information to generate a map containing only static objects during mapping. Ultimately, this method enables accurate positioning and mapping in unknown dynamic environments. The present invention also combines two geometric errors to model feature points, enabling accurate estimation of their motion state. Furthermore, a dynamic region detection algorithm based on multi-mask fusion is employed to accurately identify moving regions and generate a map containing only static objects. These innovations significantly improve the system's robustness in dynamic scenarios, enabling robots to accurately locate and map objects in these conditions. Therefore, the method presented in this paper has significant practical application value and can be widely applied to fields such as robotic navigation and autonomous driving, providing strong support for technological innovation and progress in related industries.

[0116] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A visual SLAM method based on multi-mask fusion, characterized in that: The following steps are involved: S1. Obtain an RGB image and its corresponding depth image, extract feature points from the RGB image, and obtain depth information from the depth image; S2. Classify the objects in the RGB image into C1, C2 and C3 respectively; C1 represents an absolutely stationary object, C2 represents a potentially moving object, and C3 represents a pre-set moving object; S3, use only the feature points in the C1 area to perform a preliminary estimate of the camera pose and obtain T init , according to T init Calculate the epipolar constraint error e of the feature point epi and the depth projection error e depth , the two errors are fused, and the static feature points belonging to the C1 category are modeled, and then the chi-square test is used to determine the motion status of the remaining feature points; S4. Calculate the static weights of feature points that are not dynamic points Used to weight the camera pose estimation to obtain the final camera pose estimation result T final ; S5. Calculate the dynamic area mask M based on the dynamic points g and the dynamic region mask M of the C3 category region s Fusion is performed to obtain the first dynamic area mask Based on the dynamic area detection result, the static scene is reconstructed.

2. A visual SLAM method based on multi-mask fusion according to claim 1, characterized in that, The ORB method is used to extract feature points in the RGB image and match them with the feature points in the previous key frame for camera pose estimation.

3. A visual SLAM method based on multi-mask fusion according to claim 1, characterized in that, The objects in the RGB image are classified into C1, C2 and C3, including: The RGB image is preprocessed using a semantic segmentation network, and the objects in the image are classified into C1, C2, and C3 categories according to the labels of the semantic segmentation results.

4. A visual SLAM method based on multi-mask fusion according to claim 1, characterized in that, According to T init Calculate the epipolar constraint error e of the feature point epi and the depth projection error e depth , the two errors are fused and the static feature points belonging to the C1 category are modeled. Then the chi-square test is used to determine the motion state of the remaining feature points, including: Assume that the previous key frame and the current frame are and p1 and p2 are the matching points between the two frames, where p1 is the current frame A feature point in , p2 is the previous key frame A feature point in the , the homogeneous coordinates of these two points are p1 = [u1, v1, 1] T and p2 = [u2, v2, 1] T , then the epipolar constraint error of point p1 is: Among them, u1, v1, u2, v2 represent the coordinates of the feature point in the image coordinate system, X, Y are the vectors of the polar line of the current frame, F represents the basic matrix, and the superscript T represents the transpose; project the three-dimensional point P corresponding to the feature point p1 to the previous frame Get point p' and get its projected depth value z proj , point p′ in the frame The depth value in is z′; then the depth projection error of point p1 is: yes depth =|with proj -z′| According to the preliminary estimated pose result T init , the average epipolar constraint error and depth projection error of the prior static points are calculated to be and And these two errors are independent and identically distributed, then the error distribution that the feature point in the static state should obey is: Among them, Σ is the covariance matrix, which represents the uncertainty of the scale. The calculation formula of the covariance matrix Σ of a feature point is: Among them, s represents the scale of each layer of the image pyramid extracted by the feature point, p represents the standard deviation of each layer of the image pyramid, and n represents the number of image pyramid layers where the feature point is located; the feature point polar constraint error e epi and the depth projection error e depth Combined into an error vector E = [e epi ,e depth ], for stationary feature points, the error vector composed of its errors should obey the Gaussian distribution, and its error vector E is expressed as: in, Represents the average error vector value of the feature points belonging to category C1. By calculating the inner product of the error vector E, the feature points p belonging to categories C2 and C3 are obtained. i Chi-square test value of : The obtained r i It should follow a chi-square distribution with 2 degrees of freedom: According to the feature point p i The chi-square test value r i Determine the motion state of the point:

5. A visual SLAM method based on multi-mask fusion according to claim 1, characterized in that, The calculation of the static weight of the feature points that are not dynamic points Used to weight the camera pose estimation to obtain the final camera pose estimation result T final ,include: For feature points p that are not dynamic points i , its static weight The calculation formula is: in Indicates the feature point p that is not a dynamic point i The static weight value, r i Indicates the feature point p that is not a dynamic point i Chi-square test value, Semantic(p i ) represents a feature point p that is not a dynamic point i The label category is added to the static weight to obtain the final camera pose estimation: Where R and t represent the rotation matrix and translation vector of the camera, respectively, m represents the number of feature points that are not dynamic points, and Π(·) represents the camera projection model.

6. A visual SLAM method based on multi-mask fusion according to claim 1, characterized in that, The dynamic area mask M is obtained by calculating the dynamic points g ,include: Only the dynamic points in the C2 category area are calculated to obtain the dynamic area mask M g .

7. A visual SLAM method based on multi-mask fusion according to claim 1, characterized in that, The dynamic area mask M is obtained by calculating the dynamic points g and the dynamic region mask M of the C3 category region s Fusion is performed to obtain the first dynamic area mask include: According to the Euclidean distance between dynamic points, the DBSCAN clustering method is used to group them and obtain the convex hull of each group to generate the motion area mask M. g and the dynamic region mask M of the C3 category region s Fusion is performed to obtain the first dynamic area mask 8. A visual SLAM method based on multi-mask fusion according to claim 1, characterized in that, The step S5 further includes: First dynamic area mask Divided into T independent masks And calculate each independent mask The average depth value d t : According to d t Find the image area whose depth value is within the first preset threshold range Will Decompose into K independent regions and calculate and The intersection rate between the two masks is determined by the intersection rate being greater than or equal to the second preset threshold. The effective mask in and will and Fusion obtains the second dynamic area mask H represents the number of valid masks.

9. A visual SLAM method based on multi-mask fusion according to claim 8, characterized in that, The first dynamic area mask Divided into T independent masks Expressed as: The calculation where each independent mask The average depth value d t , the calculation formula is: Among them, p j Indicates The jth pixel in the region, D(p j ) represents point p j The depth value of According to d t Find the image area whose depth value is within the first preset threshold range The judgment process is: Wherein, p represents any pixel in the image, and τ1 represents the first preset threshold; Will Decompose into K independent regions Expressed as: and calculate and The intersection rate between two masks is: Where P(·,·) represents the intersection rate between the two masks, and N(·,·) represents the number of pixels that intersect between the two masks; Determine by the intersection rate being greater than or equal to the second preset threshold τ2 The effective mask in q represents There are q conditions that meet the conditions area, represents one of the regions that meets the condition, 1≤z≤q; The said and Fusion obtains the second dynamic area mask Expressed as:

10. A visual SLAM method based on multi-mask fusion according to claim 1, characterized in that, The step of reconstructing the static scene based on the dynamic area detection result includes: When building a dense point cloud map, the point cloud in the dynamic area is removed to obtain a static and reliable point cloud map.

Citation Information

Patent Citations

  • Lightweight visual semantic odometer method for dynamic scene

    CN115273080A

  • Semantic vision SLAM method and system suitable for dynamic environment

    CN117392222A