A dynamic environment RGB-D visual slam method and device based on local observation consistency

Abstract

A dynamic environment RGB-D visual SLAM method and device based on local observation consistency, the method comprising: S1: performing grid-based improved adaptive K-Means clustering on a depth image; S2: performing sub-pose estimation based on the region of S1 clustering; S3: calculating the possibility that each region is dynamic through sub-pose; projecting the feature points of the current frame to the previous frame through sub-pose projection transformation, and then calculating the pixel distance between the corresponding points; S4: a probability grid model is proposed; it has the ability of time delay and space expansion, can effectively reduce the influence of unstable optimization of a certain block region, is more widely applicable, improves the stability of regional feature discrimination, and reduces adjustable parameters. The present application can detect unknown targets, and can solve the influence brought by dynamic targets without feedback of initial position estimation.

Description

Technical Field

[0001] This invention relates to the field of robot global localization, specifically a dynamic environment RGB-D visual SLAM method and apparatus based on local observation consistency. Background Technology

[0002] Visual SLAM is a technology that utilizes cameras as sensors in fields such as robotics, virtual reality, and autonomous driving. It can estimate the camera's position in 3D space and generate real-time sparse or dense maps. With advancements in computer vision, dense maps can also be transformed into semantic maps for performing more advanced tasks. Traditional SLAM methods often simplify problem descriptions based on assumptions about a static environment. However, this means that the performance of these methods can be severely impacted, or even lead to failure, in the presence of dynamic objects. Therefore, SLAM techniques for dynamic environments have emerged, known as dynamic SLAM. Currently, many researchers are attempting to combine traditional SLAM techniques with deep learning to address the interference that dynamic objects may cause. However, while this approach solves the problem of dynamic targets to some extent, it also introduces challenges such as poor real-time performance, missed detection of dynamic objects, and difficulties in model training.

[0003] To address the poor performance of visual SLAM in dynamic environments, researchers have implemented various measures. For example, patent CN1151115672A identifies preset targets using YOLOv5s and tracks feature points using sparse optical flow to obtain their motion information, ultimately determining which feature points are moving. Patent CN115393538A extracts features from image information using the GCNv2 algorithm and removes dynamic feature points using an optical flow dynamic point removal algorithm. Current methods are mostly based on deep learning; however, deep learning algorithms primarily identify preset targets. The key to dynamic visual SLAM should be the identification of dynamic regions, not just dynamic targets. Patent CN116299525A generates independent regions using spatial partitioning technology and effectively identifies dynamic regions based on their motion characteristics. This patent not only identifies dynamic regions but also separates them from pose estimation, enabling two independent threads. However, this method suffers from poor real-time performance and low camera information utilization.

[0004] Currently, robot localization in dynamic environments suffers from the following shortcomings: 1) Initial pose feedback is required to determine the true dynamic point, but current methods struggle to control the polarity of the feedback, potentially causing system failure in harsh environments. 2) Deep learning-based methods require a large amount of prior information, and the computational demands of the inference model are enormous, making them unsuitable for low-cost, real-time-critical applications. 3) The presence of dynamic objects in real-world environments is unavoidable, significantly impacting the device's localization accuracy and hindering further development of visual SLAM. Summary of the Invention

[0005] To address the aforementioned shortcomings in the prior art, this invention provides a dynamic environment RGB-D visual SLAM method and apparatus based on local observation consistency.

[0006] This invention improves positioning accuracy and robustness in dynamic real-world environments. It is an RGB-D SLAM system based on local observation consistency. This system can detect unknown targets and resolve the impact of dynamic targets without requiring feedback from initial position estimation. First, the spatial domain is divided into multiple independent and non-interfering spaces using a grid-based improved adaptive K-Means clustering method. Second, the three-dimensional space is reduced to a two-dimensional plane, and by observing the consistency of motion states in the two-dimensional plane, dynamic regions can be efficiently captured. Furthermore, a probabilistic grid model is proposed, which improves the robustness and accuracy of target detection through probabilistic iteration.

[0007] To achieve the above objectives, the first aspect of the present invention provides the following technical solution: a dynamic environment RGB-D visual SLAM method based on local observation consistency, comprising the following steps:

[0008] S1: Perform grid-based improved adaptive K-Means clustering on the depth image;

[0009] The specific steps of the clustering method include:

[0010] S11: Divide the depth image into w×h small grids and perform clustering operations on these small grids as units.

[0011] S12: The depth information of each cell is represented by the average depth values ​​of the center point of each cell and its n neighboring cells.

[0012]

[0013] n AThe value represents the number of grid cells in the N-neighborhood, including the center point. depth(*) represents the depth information of the original image. A(x,y) represents the set of grid cells in the N-neighborhood.

[0014] S13: Delete grid depth values ​​that are less than or equal to 0 or greater than 6 meters. These will be referred to as hollow grids in this article.

[0015] S14: The specific number of clusters can be obtained by comparing the effective grid (excluding the hollow grid) with the empirical value α:

[0016]

[0017] K represents the number of clusters calculated adaptively, α specifically represents how many effective grid cells an object might occupy, and N Valid Indicates the number of valid grid cells.

[0018] S15: Based on K calculated in S14, spatial clustering is performed to obtain C = {C1, C2...C...} K}

[0019] S2: Sub-pose estimation based on regions clustered in S1;

[0020] The specific steps for sub-pose estimation include:

[0021] S21: To achieve uniformity and retain as many feature points as possible, corner points were extracted uniformly from the quadtree feature points of ORB-SLAM2.

[0022] S22: Using the corner points extracted in S21 as input to the KLT sparse optical flow method, feature point matching is performed between two adjacent frames. The matching formula is as follows:

[0023] I x u+I y v+I t =0 (3)

[0024] Let I x I y I represents the image gradient of the gray level of a pixel along the x and y directions, respectively. t This represents the time gradient along t.

[0025] S23: Normalized Correlation (NCC) is used as a metric to verify whether a point is a corresponding corner. A point is considered a correct corresponding corner when the NCC score is higher than a set threshold ρ. The formula for calculating the NCC score is shown below:

[0026]

[0027] I(·) represents the grayscale value of the corresponding pixel, and A and B represent the w×h pixel blocks of the i-th frame and the (i-1)-th frame, respectively.

[0028] S24: The points filtered by NCC are projected from the current frame to the previous frame, and the pixel Euclidean distance d between the corresponding feature points is calculated. If the feature point distance between two frames exceeds the set threshold D, it is considered an incorrect match.

[0029] S25: Optimize all regions C above using ORBSLAM2 graph optimization to obtain... However, in order to maintain real-time performance, the total number of optimizations was reduced to x.

[0030] S3: Possessing a subpositional posture However, it's difficult to effectively measure which regions contain dynamic objects. Therefore, it's necessary to calculate the probability of each region being dynamic using sub-pose. A simple and efficient method is used to convert the feature points of the current frame into sub-poses. The projection is transformed to the previous frame, and then the pixel distance between corresponding points is calculated;

[0031] The method for calculating dynamic probability values ​​specifically includes:

[0032] S31: Points at greater distances contribute relatively little to the determination of dynamic feature attributes. A penalty value β is applied to the pixel distance between corresponding feature points with a depth value greater than 6 meters.

[0033] S32: A simplified strategy is employed to calculate the score. The distance is calculated only between two random feature points within each cell, ignoring feature points belonging to the same class. This effectively controls the computational time complexity while ensuring a constant relationship between the number of feature values ​​used in the score calculation and the set number of feature values.

[0034] S33: Before calculating the probability values, certain special cases need to be handled, such as when the number of feature points in a certain class is too small or when the sub-pose optimization gets stuck in a local optimum. In these cases, their scores may become uncontrollable. These special cases may affect the calculation of the probability values, so they need to be temporarily set as the mean of the non-special case region.

[0035] S34: First, calculate the score S(i). i∈1,2,..K The mean μ:

[0036]

[0037] S35: Calculate the variance σ:

[0038]

[0039] S36: To prevent special cases from affecting the upper limit of the data, it is necessary to reduce the contribution of special cases:

[0040]

[0041] θ represents the penalty percentage for contribution.

[0042] S37: Since the probability value represents the degree of dynamism, a reference bit, i.e., the absolutely static value of 0, needs to be added to the scoring container to obtain the scoring container S(i). i∈0,1,2,..K Calculate the mean μ' of the score container after adding the absolute static value:

[0043]

[0044] 'a' represents the number of scores less than the variance.

[0045] S38: The upper limit of the final dynamic degree can be expressed as λ:

[0046] λ=μ'+Mid{S} (9)

[0047] S39: S is the set of probability values. Based on λ in S38, the dynamic probability value of each class can be calculated.

[0048]

[0049] S4: Since the core idea of ​​the dynamic and static probability values ​​of a region is based on local observations, this method is prone to false positives and false negatives. Furthermore, motion is spatially continuous, and the edges of objects are also unstable. Therefore, a probabilistic grid model is proposed. It possesses the ability to accommodate time delays and spatial expansion, effectively reducing the impact of instability in optimizing a specific region, thus having wider applicability, improving the stability of region feature discrimination, and reducing the number of adjustable parameters.

[0050] The probabilistic grid model method specifically includes:

[0051] S41: Due to scene variations, it is difficult to determine whether a region is dynamic based on a fixed threshold. Therefore, this paper adopts a dynamic thresholding method to set the threshold ω:

[0052]

[0053] M(j) represents the absolute median error, which can be expressed as:

[0054] M(j)=|P(j)-Mid{P}|,j∈0,1...K (12)

[0055] P is a set of probability values, and the overall volatility of the data is calculated based on the deviation of each probability value from the median. b is a correction value.

[0056] S42: Probability updates can be based on:

[0057]

[0058] M t ∈[-1,1] is the final dynamically generated mask, also known as the target mask, because M t The probability values ​​need to be maintained and updated for each frame. P(i) is the probability function of region i being a dynamic region. misshit and hit represent the weights for hitting or missing, respectively. When M t When the value is greater than 0, it is considered a dynamic region.

[0059] S43: Utilizing the probability region crossover ratio to quickly update dynamic regions can effectively prevent lag in region recognition. This means that if a region has been considered static for a long time and a moving object suddenly enters, the system may fail to quickly recognize it as a new dynamic region. This situation can negatively impact system performance.

[0060] S341: Calculate the average probability value of the grid cells with a probability greater than zero on the target board in the i-th type region:

[0061]

[0062] n represents M t The number of grid cells where (j) > 0, c i This represents the total number of grid cells in the i-th class. If the probability value P′ mask (i) If the value of n is greater than the set value p and the proportion of n in the number of grids in the i-th type region of the current frame is greater than the threshold r, it means that the target has moved to a new region.

[0063] A second aspect of the present invention relates to a dynamic environment RGB-D visual SLAM device based on local observation consistency, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement a dynamic environment RGB-D visual SLAM method based on local observation consistency according to the present invention.

[0064] A third aspect of the invention relates to a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements a dynamic environment RGB-D visual SLAM method based on local observation consistency according to the invention.

[0065] This invention can be used as a front-end preprocessing step in the ORBSLAM2 / 3 system to filter out dynamic regions, thereby reducing erroneous data associations. This invention designs a grid-based adaptive K-Means clustering algorithm, which adaptively adjusts the number of clusters based on the effective number of grid cells, effectively solving the problem of large runtime variance caused by clustering. Furthermore, a dynamic region identification method based on local observation consistency is proposed. It utilizes nonlinear least squares optimization based on reprojection error to obtain the approximate pose of each region, i.e., the sub-pose, thereby quantifying the dynamic probability of each region. This method features independent identification, requiring no prior correction of image distortion or model training, while efficiently and accurately detecting dynamic regions. Finally, this invention designs a probability-based grid model to enhance the accuracy and robustness of dynamic region identification. This invention effectively reduces the impact of dynamic objects on robot localization, improving the system's robustness and scene generalization.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] This invention presents a dynamic environment RGB-D visual SLAM method based on local observation consistency. It addresses the limitation of traditional dynamic visual SLAM, which requires initial pose estimation feedback to discriminate dynamic regions. Furthermore, it achieves this without the need for prior knowledge of training samples and dynamic objects. Simultaneously, it improves not only the localization accuracy and robustness of visual SLAM in dynamic environments but also its real-time performance. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method of the present invention;

[0069] Figures 2a to 2b This is a schematic diagram illustrating the effect of adaptive clustering as described in the embodiments of the present invention, wherein, Figure 2a It is the original image. Figure 2b This is a schematic diagram of the clustering results. Different colors represent different clusters, and each circle represents the cluster to which the rectangle belongs.

[0070] Figures 3a to 3b This is a schematic diagram illustrating the detection effect of the dynamic region as described in the embodiments of the present invention, wherein, Figure 3a This is the original image, where the green dots are used as feature points for tracking. Figure 3b This is a schematic diagram of the dynamic region detection results;

[0071] Figures 4a to 4b This is a comparison image of the present invention and the actual trajectory on the TUM dataset. Figure 4a The camera is translating along the XYZ axes. Figure 4b The camera moves along a hemisphere with a radius of 1 meter. The shorter the red line, the smaller the positioning error.

[0072] Figures 5a to 5b This is a schematic diagram illustrating the dynamic region detection effect in a real-world scenario as described in the embodiments of the present invention. Figure 5a This is the original image, where the green dots are used as feature points for tracking. Figure 5b This is a schematic diagram of the dynamic region detection results. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the described embodiments are only for explaining the invention and are not intended to limit the invention.

[0074] Example 1

[0075] This invention is a dynamic environment RGB-D visual SLAM method based on local observation consistency, comprising the following steps:

[0076] S1: Perform grid-based improved adaptive K-Means clustering on the depth image;

[0077] The specific steps of the clustering method include:

[0078] S11: Divide the depth image into 30×25 small grids and perform clustering operations on these small grids.

[0079] S12: The depth information of each cell is represented by the average depth values ​​of its center point and its n=4 adjacent cells.

[0080]

[0081] n A The value represents the number of grid cells in the N-neighborhood, including the center point. depth(*) represents the depth information of the original image. A(x,y) represents the set of grid cells in the N-neighborhood.

[0082] S13: Delete grid depth values ​​that are less than or equal to 0 or greater than 6 meters. These will be referred to as hollow grids in this article.

[0083] S14: The specific number of clusters can be obtained by comparing the effective grid (excluding the hollow grid) with the empirical value α = 75.

[0084]

[0085] K represents the number of clusters calculated adaptively, α specifically represents how many effective grid cells an object might occupy, and N Valid Indicates the number of valid grid cells.

[0086] S15: Based on K calculated in S14, spatial clustering is performed to obtain C = {C1, C2...C...} K}

[0087] S2: Sub-pose estimation based on regions clustered in S1;

[0088] The specific steps for sub-pose estimation include:

[0089] S21: To achieve uniformity and retain as many feature points as possible, corner points were extracted uniformly from the quadtree feature points of ORB-SLAM2.

[0090] S22: Using the corner points extracted in S21 as input to the KLT sparse optical flow method, feature point matching is performed between two adjacent frames. The matching formula is as follows:

[0091] I x u+I y v+I t =0 (3)

[0092] Let I x I y I represents the image gradient of the gray level of a pixel along the x and y directions, respectively. t This represents the time gradient along t.

[0093] S23: Normalized Correlation (NCC) is used as a metric to verify whether a point is a corresponding corner. A point is considered a correct corresponding corner when the NCC score is higher than the set threshold ρ = 0.8. The formula for calculating the NCC score is as follows:

[0094]

[0095] I(·) represents the grayscale value of the corresponding pixel, and A and B represent the w×h pixel blocks of the i-th frame and the (i-1)-th frame, respectively.

[0096] S24: The points filtered by NCC are projected from the current frame to the previous frame, and the pixel Euclidean distance d between the corresponding feature points is calculated. If the feature point distance between two frames exceeds the set threshold D=25, it is considered an incorrect match.

[0097] S25: Optimize all regions C above using ORBSLAM2 graph optimization to obtain... However, in order to maintain real-time performance, the total number of optimizations was reduced to x = 9.

[0098] S3: Possessing a subpositional posture However, it's difficult to effectively measure which regions contain dynamic objects. Therefore, it's necessary to calculate the probability of each region being dynamic using sub-pose. A simple and efficient method is used to convert the feature points of the current frame into sub-poses. The projection is transformed to the previous frame, and then the pixel distance between corresponding points is calculated;

[0099] The method for calculating dynamic probability values ​​specifically includes:

[0100] S31: Points at a distance contribute relatively little to the determination of dynamic feature attributes. A penalty value β = 0.45 is applied to the pixel distance between corresponding feature points with a depth value greater than 6 meters.

[0101] S32: A simplified strategy is employed to calculate the score. The distance is calculated only between two random feature points within each cell, ignoring feature points belonging to the same class. This effectively controls the computational time complexity while ensuring a constant relationship between the number of feature values ​​used in the score calculation and the set number of feature values.

[0102] S33: Before calculating the probability values, certain special cases need to be handled, such as when the number of feature points of a certain class is too small or when the sub-pose optimization gets stuck in a local optimum. In these cases, their scores may become uncontrollable. These special cases may affect the calculation of the probability values, so they need to be temporarily set as the mean of the effective region.

[0103] S34: First, calculate the score S(i). i∈1,2,..K The mean μ:

[0104]

[0105] S35: Calculate the variance σ:

[0106]

[0107] S36: To prevent special cases from affecting the upper limit of the data, it is necessary to reduce the contribution of special cases:

[0108]

[0109] θ = 0.5 represents the penalty ratio for contribution.

[0110] S37: Since the probability value represents the degree of dynamism, a reference bit, i.e., the absolutely static value of 0, needs to be added to the scoring container to obtain the scoring container S(i). i∈0,1,2,..K Calculate the mean μ' of the score container after adding the absolute static value:

[0111]

[0112] 'a' represents the number of scores less than the variance.

[0113] S38: The upper limit of the final dynamic degree can be expressed as λ:

[0114] λ=μ'+Mid{S} (9)

[0115] S39: S is the set of probability values. Based on λ in S38, the dynamic probability value of each class can be calculated.

[0116]

[0117] S4: Since the core idea of ​​the dynamic and static probability values ​​of a region is based on local observations, this method is prone to false positives and false negatives. Furthermore, motion is spatially continuous, and the edges of objects are also unstable. Therefore, a probabilistic grid model is proposed. It possesses the ability to accommodate time delays and spatial expansion, effectively reducing the impact of instability in optimizing a specific region, thus having wider applicability, improving the stability of region feature discrimination, and reducing the number of adjustable parameters.

[0118] The probabilistic grid model method specifically includes:

[0119] S41: Due to scene variations, it is difficult to determine whether a region is dynamic based on a fixed threshold. Therefore, this paper adopts a dynamic thresholding method to set the threshold ω:

[0120]

[0121] M(j) represents the absolute median error, which can be expressed as:

[0122] M(j)=|P(j)-Mid{P}|,j∈0,1...K (12)

[0123] P is a set of probability values. The overall volatility of the data is calculated based on the deviation of each probability value from the median. b = 0.5 is the correction value.

[0124] S42: Probability updates can be based on:

[0125]

[0126] M t ∈[-1,1] is the final dynamically generated mask, also known as the target mask, because M t The probability value needs to be maintained and updated for each frame. P(i) is the probability function that region i is a dynamic region.

[0127] misshit = 0.57 and hit = -0.37 represent the weights for hitting or missing, respectively.

[0128] S43: Utilizing the probability region crossover ratio to quickly update dynamic regions can effectively prevent lag in region recognition. This means that if a region has been considered static for a long time and a moving object suddenly enters, the system may fail to quickly recognize it as a new dynamic region. This situation can negatively impact system performance.

[0129] S341: Calculate the average probability value of the grid cells with a probability greater than zero on the target board in the i-th type region:

[0130]

[0131] n represents M t The number of grid cells where (j) > 0, c i This represents the total number of grid cells in the i-th class. If the probability value P′ mask (i) If the value of n is greater than the set value p = 0.75 and the proportion of n in the number of grid cells in the i-type region of the current frame is greater than the threshold r = 0.35, it means that the target has moved to a new region.

[0132] As shown in Figure 2, without needing to set the number of clusters, related objects can be adaptively clustered into a single cluster, improving the system's robustness. Figure 3 shows the identification of dynamic regions on the TUM dataset. This invention effectively identifies dynamic regions, eliminating interference from dynamic objects on the positioning system. Figure 4 is a comparison of the absolute positioning accuracy of this invention with the actual trajectory. The shorter the red line, the smaller the positioning error. The trajectory of this invention closely follows the actual trajectory, proving that this invention can improve the positioning accuracy and robustness of the SLAM system in dynamic environments. Appendix Figure 5 is a visualization of dynamic region identification in a real-world environment, fully demonstrating the reliability and practicality of this invention.

[0133] Traditional visual SLAM, due to its assumption of a static environment, is often affected by the presence of dynamic objects, impacting its performance. Recent works have attempted to eliminate the influence of dynamic targets; however, these methods are limited by initial position estimation (based on feedback estimation) and unknown targets. This embodiment proposes an RGB-D SLAM system based on local observation consistency, which can detect unknown targets and resolve the influence of dynamic targets without the need for feedback from initial position estimation. First, the spatial region is divided into multiple independent and non-interfering spaces using a grid-based improved adaptive K-Means clustering method. Second, the dimensionality of the three-dimensional space is reduced to a two-dimensional plane; by observing the consistency of motion states in the two-dimensional plane, dynamic regions can be efficiently captured. Furthermore, this invention proposes a probabilistic grid model, which improves the robustness and accuracy of target detection through probabilistic iteration. This invention is experimentally evaluated on the challenging TUM dynamic sequence dataset and in real-world environments.

[0134] Example 2

[0135] This embodiment relates to a dynamic environment RGB-D visual SLAM device based on local observation consistency, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a dynamic environment RGB-D visual SLAM method based on local observation consistency according to Embodiment 1.

[0136] Example 3

[0137] This embodiment relates to a computer-readable storage medium storing a program that, when executed by a processor, implements a dynamic environment RGB-D visual SLAM method based on local observation consistency according to Embodiment 1.

[0138] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation schemes. Any other similar implementations derived by those skilled in the art based on the technical solutions of this invention also fall within the protection scope of this invention.

Claims

1. A dynamic environment RGB-D visual SLAM method based on local observation consistency, characterized in that, Includes the following steps: S1: Perform grid-based improved adaptive K-Means clustering on the depth image; S2: Sub-pose estimation based on regions clustered in S1; S3: Calculate the probability that each region is dynamic through sub-pose, and apply the feature points of the current frame to the sub-pose. The projection is transformed to the previous frame, and then the pixel distance between corresponding points is calculated; S4: Propose a probabilistic grid model; the probabilistic grid model includes: S41: Use a dynamic threshold method to set the threshold. : （11） The absolute median error is expressed as: （12） It is a set of probability values, and the overall degree of data fluctuation is calculated based on the deviation of each probability value from the median. This is a correction value; S42: Probability update based on: （13） The final dynamically generated mask is also called the target mask, because... The probability value needs to be maintained and updated for each frame; It is the probability function that region i is a dynamic region; These are respectively represented as the weights for hitting or missing; when When the value is greater than 0, it is considered a dynamic region; S43: By using the probability region cross ratio to quickly update the dynamic region, it can effectively prevent the lag phenomenon in region recognition; S431: Calculate the average probability value of the grid cells with a probability greater than zero on the target board in the i-th type region: （14） express The number of grid cells, Represents the total number of grid cells in the i-th class; if the probability value Greater than the set value ,and The proportion of the current frame's grid cells in region i is greater than the threshold. This means that the target has moved to a new area.

2. The dynamic environment RGB-D visual SLAM method based on local observation consistency as described in claim 1, characterized in that, The clustering described in step S1 includes: S11: Divide the depth image into w×h small grids and perform clustering operations on these small grids as units; S12: This uses the center point of each grid and its... The depth information of a cell is represented by the average depth values ​​of its neighboring cells. （1） This represents the number of cells in the N-squared area, including the center point. Represents the depth information of the original image. A set of lattices representing the neighborhood of N; S13: Delete grid depth values ​​that are less than or equal to 0 or greater than 6 meters, which are called hollow grids; S14: Based on valid meshes other than hollow meshes and empirical values The specific number of clusters can be obtained by taking the ratio: （2） This represents the number of clusters calculated adaptively. Specifically, it means how many effective grid cells an object might occupy. Indicates the number of valid grid cells; S15: Spatial clustering is performed based on K calculated in S14 to obtain... .

3. The dynamic environment RGB-D visual SLAM method based on local observation consistency as described in claim 1, characterized in that, Step S2, which involves sub-pose estimation based on the region clustered in S1, includes: S21: Corner points uniformly extracted using the quadtree feature points of ORB-SLAM2; S22: Using the corner points extracted in S21 as input to the KLT sparse optical flow method, feature point matching is performed between two adjacent frames. The matching formula is as follows: （3） set up , These represent the gray levels of pixels in the image along... , Image gradient in direction, Indicates along Time gradient; S23: Use Normalized Correlation (NCC) as a metric to verify whether it is a corresponding corner point; when the score of Normalized Correlation is higher than the set threshold... When the corresponding corner point is selected, it is considered the correct corner point; the formula for calculating the normalized correlation score is as follows: （4） Represents the grayscale value of the corresponding pixel, and A and B represent the w×h pixel blocks in the i-th frame and the (i-1)-th frame, respectively; S24: The points filtered by NCC are projected from the current frame to the previous frame, and the pixel Euclidean distance d between the corresponding feature points is calculated. If the feature point distance between two frames exceeds the set threshold D, it is considered an incorrect match. S25: Utilize ORBSLAM2 graph optimization for all the aforementioned regions. Optimization results However, in order to maintain real-time performance, the total number of optimizations was reduced to [number missing]. Second-rate.

4. The dynamic environment RGB-D visual SLAM method based on local observation consistency as described in claim 1, characterized in that, Step S3, which involves calculating the probability that each region is dynamic through sub-pose, includes: S31: Distant points contribute relatively little to the determination of dynamic feature attributes; a penalty value is applied to the pixel distance between corresponding feature points with a depth value greater than 6 meters. ; S32: Calculate the distance between only two random feature points within each cell, without considering feature points belonging to the same class; S33: Before calculating the probability value, for special cases where the number of feature points is too small or the sub-pose optimization is trapped in a local optimum, the special case is temporarily set as the mean of the effective region. ; S34: First calculate the score. mean : （5） S35: Calculate Variance : （6） S36: To prevent special cases from affecting the upper limit of the data, the contribution of special cases is reduced: （7） Indicates the penalty percentage for contribution; S37: Since the probability value represents the degree of dynamism, adding a reference bit (i.e., the absolutely static value of 0) to the scoring container yields the scoring container. Calculate the mean of the score container after adding absolute static values. : （8） This indicates the number of scores less than the variance. S38: The upper limit of the final dynamic level can be expressed as : （9） S39: S is the set of probability values, according to S38 Calculate the dynamic probability value for each class: （10）。 5. A dynamic environment RGB-D visual SLAM device based on local observation consistency, characterized in that, The system includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement a dynamic environment RGB-D visual SLAM method based on local observation consistency as described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements a dynamic environment RGB-D visual SLAM method based on local observation consistency as described in any one of claims 1-4.

Images