A SLAM method and system for dense point clouds in a dynamic environment based on YOLOv11 and ORB-SLAM3

Abstract

The present invention discloses a SLAM method and system for dense point clouds in a dynamic environment based on YOLOv11 and ORB-SLAM3. This method integrates the real-time object detection and image segmentation technologies of the YOLOv11 model into the ORB-SLAM3 framework, achieving high-precision and robust visual SLAM in a dynamic environment. By using the balanced convolutional method GSConv layer to replace the traditional convolutional layer in the YOLOv11 model and adopting the new feature fusion module VoVGSCS layer to replace the traditional C2f layer, the Neck structure of YOLOv11 is improved, and a lightweight network model is achieved. Experimental data confirm that the pose estimation accuracy of this method in a dynamic environment is significantly better than that of existing visual SLAM algorithms.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent robots, and particularly relates to a SLAM method and system for dense point clouds in a dynamic environment based on YOLOv11 and ORB-SLAM3. Background Art

[0002] With the development of society and the progress of technology, intelligent robots have been widely used in various fields. As one of the key technologies in the field of intelligent robot research, SLAM technology is crucial for the autonomous navigation of intelligent robots. However, achieving accurate and robust visual SLAM in a dynamic environment remains a major challenge. The present invention aims to propose a method based on improved YOLOv11 fused with ORB-SLAM3 to address the dense point cloud SLAM problem in a dynamic environment. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention provides a SLAM method and system for dense point clouds in a dynamic environment based on YOLOv11 and ORB-SLAM3. The system includes three modules: feature extraction, feature matching, and feature point elimination. The technical solution of the present invention is realized through the following content:

[0004] The first aspect of the present invention: A SLAM method for dense point clouds in a dynamic environment based on the YOLOv11 model and ORB-SLAM3, the method comprising the following steps:

[0005] (1) Capturing an image by a structured light system in a dynamic environment, and performing dynamic target detection and image segmentation based on the improved YOLOv11 model;

[0006] (2) According to the detection result, determining whether the dynamic feature points are located within the detection frame of the dynamic object. If they are located within the detection frame of the dynamic object, the dynamic feature points are eliminated;

[0007] (3) Inputting the remaining static feature points into the SLAM system for pose estimation and map construction; and evaluating and optimizing the system robustness according to the pose estimation result;

[0008] (4) Verifying the dynamic feature points of consecutive frames by using temporal consistency checking and further weighted elimination to complete pose estimation and mapping.

[0009] Specifically, the improved YOLOv11 model includes an input layer, a preprocessing layer, multiple standard convolutional layers, a GSConv layer, a VoVGSCSP feature fusion layer, and an output layer. The GSConv layer is used to replace the traditional convolutional layer to balance accuracy and computational load, and the VoVGSCSP feature fusion layer is used to replace the C2f feature fusion layer.

[0010] Furthermore, step (1) includes the following steps:

[0011] (1.1) Multi-scale feature extraction

[0012] Extract feature maps according to different levels of the input captured by the structured light system: ;

[0013] wherein, , are the height, width, and number of channels of the feature map respectively;

[0014] (1.2) Spatial pyramid pooling

[0015] Aggregate the multi-scale feature extraction results of the input layer through pooling operations of different scales,

[0016] Output multi-scale features: , For all within a window of size ; where represents a pooling operation with a kernel size of ;

[0017] Process the SPP output features through the GSConv convolutional layer. GSConv combines standard convolution and depthwise separable convolution, and the weight parameter is , and the output is: ;

[0018] Activation function : ;

[0019] Final fused output features: ;

[0020] (1.3) Input layer

[0021] Send the spatial pyramid pooling result, after normalization, into the network for forward propagation;

[0022] Adjust the input image to the standard size required by the YOLOv11 model and normalize it to adapt to the model input requirements; the normalization formula is: ; where is the pixel value of the original image, is the pixel value after normalization;

[0023] (1.4) Object detection

[0024] Use the model to generate dynamic object detection results, including object class labels, confidence levels, and bounding boxes, where each bounding box is defined as The dynamic object in the model; the data flow of the model first convolves the multi-scale feature image, reduces the number of channels of the original C1 to C2, and names it T1;

[0025] Then, T1 is subjected to a depth-wise separable convolution operation to extract deep features, and is named T2. Finally, T1 and T2 are merged to form a result with twice the number of channels of C2, named T3. Finally, T3 is reshuffled according to the eigenvalues:

[0026] FSPP=concat(poolk1(Ft),poolk2(Ft),…,poolkn(Ft));

[0027] The output of the dynamic target detection results generated by the model includes the following information:

[0028] (a) Target category label c, indicating the category of the detected target;

[0029] (b) Confidence s, indicating the probability that the target belongs to category c;

[0030] (c) Bounding box Circumference, of which:

[0031] : Coordinates of the upper left corner of the bounding box; : The coordinates of the lower right corner of the bounding box; combine the results and express it as: ; Where i is the index of the target and N is the total number of targets detected in the current frame.

[0032] Furthermore, the step (2) specifically includes the following steps:

[0033] (2.1) Use the ORB algorithm to extract feature points and their descriptors from the current frame. The feature points are represented by pixel coordinates. express;

[0034] (2.2) For each dynamic target detected in step (1), obtain its bounding box coordinates and mark all feature points within the bounding box as dynamic feature points; the determination conditions for dynamic feature points are as follows: ;when If the above conditions are met, it is determined to be a dynamic point, otherwise it is retained as a static point;

[0035] (2.3) For the set of static feature points after elimination, retain their coordinates and descriptors for subsequent processing of the SLAM system.

[0036] Furthermore, the step (3) specifically includes the following steps:

[0037] (3.1) Pose Estimation

[0038] Perform a preliminary estimation of the camera pose based on static feature points; ORB-SLAM3 calculates the camera pose through feature point matching and triangulation methods:

[0039] For the set of static feature points: ;

[0040] And the set of feature points in the key frame: ;

[0041] Perform descriptor matching, and the expression for calculating the similarity between two points using the Hamming distance is as follows: ; where represents the exclusive OR operation, is a function for counting the number of 1s, and retain the matching pairs with the smallest distance and lower than the threshold ; ;

[0042] (3.2) Pose estimation between two frames

[0043] Perform robust estimation through two models, the fundamental matrix F and the homography matrix H, and use the RANSAC algorithm to filter out the mismatched point pairs to obtain a set of matching points that meet the geometric constraints; the fundamental matrix is used to describe the epipolar geometry relationship of point pairs in two images. For the matching point pair , it satisfies:

[0044] ; where and represent the homogeneous coordinates of the same scene point in the two images;

[0045] The homography matrix describes the perspective transformation between two images and is used for the case of pure rotation or planar scenes; for the matching point pair , it satisfies: ; that is: ; where is a scale factor;

[0046] (3.3) Optimize pose estimation

[0047] Estimate the pose of the current frame through the PnP problem according to the set of matching points , and its optimization goal is to minimize the reprojection error: ;

[0048] where, is the feature point in the image coordinate system; is the feature point in three-dimensional space, K is the camera intrinsic matrix, is the camera extrinsic parameter, is the projection function: ;

[0049] Solve the minimum value of through the Gauss-Newton method or the LM algorithm to obtain the pose ;

[0050] (3.4) Optimize the map and pose

[0051] Construct a graph optimization-based model, optimize the map and pose by minimizing the error cost function, and eliminate the abnormal errors caused by residual dynamic points. In global optimization, use the graph-based optimization method to jointly optimize the pose and the map. The goal is to minimize the error cost function of the entire graph model: ; where is the observation error, is the observed value, is the estimated value, is the observation error covariance matrix;

[0052] Solve the above cost function through the G2O tool or Ceres Solver to optimize the node pose and the map point position.

[0053] Furthermore, step (3.2) further includes using the RANSAC algorithm to achieve robust model estimation by performing model fitting on randomly sampled point pairs. The specific steps are as follows:

[0054] (a) Randomly extract the minimum point set from the set of matching points for calculating the candidate fundamental matrix or the homography matrix ;

[0055] (b) Use the fitting model to evaluate the geometric error of all matching points :

[0056] For the fundamental matrix: ;

[0057] The point pairs that satisfy are inliers; where is the manually input threshold;

[0058] For the homography matrix: ;

[0059] The point pairs that satisfy are inliers;

[0060] (c) Calculate the number of inliers and select the model with the most inliers as the final estimation result;

[0061] Through the above steps, mis-matched point pairs are filtered out, and a set of point pairs that satisfy geometric constraints is retained: 。

[0062] Furthermore, step (4) includes the following steps:

[0063] (4.1) Temporal sequence consistency check

[0064] Due to the consistency of detecting dynamic point changes in consecutive frames; when a certain feature point is marked as a dynamic point in multiple consecutive frames, a rejection strategy is applied to this area. Specifically: the dynamic feature points in consecutive frames are verified using temporal consistency check, as follows:

[0065] First, define the dynamic state sequence of feature points in consecutive frames: ;

[0066] where represents that the feature point is marked as a dynamic point in the t-th frame, represents a static point;

[0067] Second, for each feature point, calculate its temporal consistency score: ;

[0068] where T is the temporal window size; when , is the dynamic point threshold, then it is considered that is a dynamic point and is rejected, otherwise it is retained as a static point;

[0069] Finally, perform a secondary verification on the feature points with abnormal temporal consistency check results, and determine whether to reject them by statistically judging the consistency of the movement directions of point trajectories: ; ; if , then reject this point;

[0070] (4.2) Weighted rejection: Assign weights to the rejection of feature points according to the confidence of the YOLOv11 detection box; the feature points within the detection box with high confidence are preferentially rejected, while the feature points in the low-confidence area are further judged.

[0071] The second aspect of the present invention: A SLAM system for dense point clouds in a dynamic environment based on the YOLOv11 model and ORB-SLAM3, the system includes the following modules:

[0072] Dynamic object detection and segmentation module: In a structured light system in a dynamic environment, images are captured, and dynamic object detection and image segmentation are performed based on the improved YOLOv11 model;

[0073] Dynamic Feature Point Determination and Elimination Module: According to the detection results, determine whether the dynamic feature points are within the detection frame of the dynamic object. If they are within the detection frame of the dynamic object, then eliminate the dynamic feature points;

[0074] Pose Estimation and Map Construction Module: Input the remaining static feature points into the SLAM system for pose estimation and map construction; and evaluate and optimize the system robustness according to the pose estimation results;

[0075] Weighted Dynamic Feature Point Elimination Module: Use temporal consistency check to verify the dynamic feature points in consecutive frames and further eliminate them by weighting to complete pose estimation and mapping.

[0076] The third aspect of the present invention: An electronic device, comprising:

[0077] One or more processors;

[0078] A memory for storing one or more programs;

[0079] When the one or more programs are executed by the one or more processors, the one or more processors implement the SLAM method for dense point clouds in a dynamic environment based on the YOLOv11 model and ORB - SLAM3.

[0080] The fourth aspect of the present invention: A computer - readable storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the SLAM method for dense point clouds in a dynamic environment based on the YOLOv11 model and ORB - SLAM3 are implemented.

[0081] The beneficial effects of the present invention are as follows:

[0082] This paper utilizes the high - precision object detection ability of the YOLOv11 model to eliminate dynamic feature points from the scene, thus effectively reducing the interference of dynamic objects on the visual SLAM system. By introducing this elimination method in the visual odometry stage of ORB - SLAM3, the problem of false matching caused by moving objects in a dynamic scene is solved, effectively improving the positioning accuracy and trajectory stability of the system. To better meet the requirements of a real - time SLAM system, this paper makes targeted improvements to YOLOv11. Specifically, the GSConv convolution technology and the VoVGSCSP feature fusion layer are introduced, significantly improving the lightweight level and inference speed of the model, while maintaining a high detection accuracy in dynamic object detection tasks. Experimental results show that the improved YOLOv11 effectively reduces the computational complexity of the system while maintaining the detection performance, laying a foundation for real - time dynamic point elimination. Description of the Drawings

[0083] The attached drawings are incorporated into the specification and are an integral part of the specification, explaining the implementation manners in which each module involved in the present invention follows the principles of the present invention. The focus of the drawings is not on limitation but on explaining the principles of the invention. In the drawings,

[0084] Figure 1 is the execution flowchart of each module of the dynamic feature point elimination system of the present invention;

[0085] Figure 2 is the data flowchart of the GSConv convolutional layer of the present invention;

[0086] Figure 3 is the epipolar geometry structure diagram in the pose estimation process between two frames of the present invention;

[0087] Figure 4 is the schematic structural diagram of the system of the present invention. Specific Embodiments

[0088] The following detailed description relates to the aforesaid drawings and will set forth the specific details of an embodiment in order to provide a comprehensive understanding of various aspects of the claimed invention. For developers familiar with the related art, it is obvious that other embodiments different from the following specific details can be adopted to implement certain modules and systems in the present invention. The following description is for the purpose of explanation rather than limitation. All modifications, equivalent substitutions, etc. made within the spirit and principles of the invention shall be included within the protection scope of the invention.

[0089] According to the solution of the present invention, the improved YOLOv11 and the integrated ORB-SLAM3 system modules and processes are designed as Figure 1 shown. The present invention applies this solution to a structured light system. The specific solution is as follows:

[0090] S1: Preprocessing of the input image and target detection

[0091] (1) Image preprocessing

[0092] (1.1) Multi-scale feature extraction

[0093] Extract feature maps according to different levels of the input captured by the structured light system: ; where , are the height, width, and number of channels of the feature map respectively.

[0094] (1.2) Spatial pyramid pooling:

[0095] Aggregate the multi-scale feature extraction result map of the input layer through pooling operations of different scales to output multi-scale features ; For all within a window of size ; where represents a pooling operation with a kernel size of .

[0096] The SPP output features are processed through the GSConv convolutional layer. GSConv combines standard convolution and depthwise separable convolution, with weight parameters of , and the output is: ; Activation function : ; Final fused output features: ;

[0097] (1.3) Input layer

[0098] The spatial pyramid pooling result is normalized and then fed into the network for forward propagation;

[0099] The input image is resized to the standard size required by the YOLOv11 model and normalized to fit the model input requirements; the normalization formula is: ; where is the pixel value of the original image, is the normalized pixel value.

[0100] (1.4) Object detection

[0101] The model is used to generate dynamic object detection results, including object class labels, confidence levels, and bounding boxes (Bounding Box), where each bounding box is defined as for the dynamic objects in

[0102] Among them, the data flow of the model is as shown in Figure 2 : First, the multi-scale feature image undergoes a convolution operation to reduce the original number of C1 channels to C2 and is named T1;

[0103] Then, T1 undergoes a depthwise separable convolution operation to extract depth features and is named T2;

[0104] Finally, T1 and T2 are merged to form a result with twice the number of C2 channels, named T3;

[0105] Finally, T3 is reshuffled according to the eigenvalues: FSPP = concat(poolk1(Ft), poolk2(Ft), …, poolkn(Ft)) ;

[0106] The output of the model for generating dynamic object detection results includes the following information:

[0107] (a) Object class label c, representing the class of the detected object.

[0108] (b) Confidence s, representing the probability that the target belongs to class c.

[0109] (c) Bounding box surrounded by, where:

[0110] : Coordinates of the upper left corner of the bounding box.

[0111] : Coordinates of the lower right corner of the bounding box.

[0112] Combining these results, it can be expressed as: ;

[0113] where: i is the index of the target, and N is the total number of targets detected in the current frame.

[0114] S2: Identification and elimination of dynamic feature points

[0115] According to the target detection results of step S1, the dynamic regions in the image are matched one by one with the feature points extracted by ORB-SLAM3, and the feature points falling within the dynamic regions are eliminated. The specific steps are as follows:

[0116] (2.1) Use the ORB algorithm to extract feature points and their descriptors from the current frame. The feature points are represented by pixel coordinates .

[0117] (2.2) For each dynamic target detected in S1, obtain its bounding box coordinates and mark all feature points within the bounding box as dynamic feature points. The determination conditions for dynamic feature points are as follows: ; If the above conditions are met, it is determined as a dynamic point; otherwise, it is retained as a static point.

[0118] (2.3) For the set of static feature points after elimination, retain their coordinates and descriptors for subsequent processing of the SLAM system.

[0119] S3: Input of static feature points and SLAM processing

[0120] Input the static feature points after eliminating dynamic points into ORB-SLAM3 to complete the following operations:

[0121] (3.1) Pose estimation

[0122] Based on the static feature points, perform a preliminary estimation of the camera pose. ORB-SLAM3 calculates the camera pose through feature point matching and triangulation methods: for the set of static feature points: ; and the set of feature points in the key frame: ; Perform descriptor matching and calculate the similarity between two points using the Hamming distance: ; where represents the exclusive OR operation, is a function to count the number of 1s. Retain the matching pairs with the smallest distance and below the threshold . .

[0123] (3.2) Pose estimation between two frames

[0124] Perform robust estimation through two models: the fundamental matrix F and the homography matrix H. Use the RANSAC algorithm to filter out mis-matched point pairs and obtain a set of matching points that satisfy geometric constraints:

[0125] Fundamental matrix is used to describe the epipolar geometry relationship between point pairs in two images. The epipolar relationship is as Figure 3 shown. For the matching point pair , it satisfies: where and represent the homogeneous coordinates of the same scene point in the two images.

[0126] Homography matrix describes the perspective transformation between two images and is applicable to the case of pure rotation or planar scenes. For the matching point pair , it satisfies: ;

[0127] That is: ; where is a scale factor.

[0128] To achieve robust model estimation, use the RANSAC algorithm to fit the model by randomly sampling point pairs. The specific steps are as follows:

[0129] Randomly extract the minimum point set from the set of matching points to calculate the candidate fundamental matrix F F or the homography matrix H H.

[0130] Evaluate the geometric error of all matching points using the fitted model:

[0131] For the fundamental matrix:

[0132] Points that satisfy are inliers.

[0133] For the homography matrix ;

[0134] Points that satisfy Pairs of points are inliers.

[0135] Calculate the number of inliers and select the model with the most inliers as the final estimation result.

[0136] Through the above steps, filter out the mismatched point pairs and retain the set of point pairs that satisfy the geometric constraints: ;

[0137] where is a manually input threshold, which is determined according to the accuracy of structured light and the stability of the inertial navigation dynamic structure.

[0138] (3.3) Optimize pose estimation

[0139] According to the set of matched points, estimate the pose of the current frame through the Perspective-n-Point (PnP) problem , and its optimization objective is to minimize the reprojection error: ;

[0140] where, is the feature point in the image coordinate system; is the feature point in 3D space, K is the camera intrinsic matrix, is the camera extrinsic parameter, is the projection function: .

[0141] Solve the minimum value of through the Gauss-Newton method or the Levenberg-Marquardt (LM) algorithm to obtain the pose .

[0142] (3.4) Optimize the map and pose

[0143] Construct a graph optimization-based model to optimize the map and pose by minimizing the error cost function, and eliminate the abnormal errors that may be caused by residual dynamic points. In global optimization, use the graph-based optimization method to jointly optimize the pose and the map, and the goal is to minimize the error cost function of the entire graph model: ; where, is the observation error, is the observation value, is the estimated value, is the observation error covariance matrix.

[0144] Solve the above cost function through the General Graph Optimization (G2O) tool or the Ceres Solver to optimize the node pose and the map point position.

[0145] S4: Improved Strategies for Performance Optimization and Dynamic Point Culling

[0146] To improve the accuracy of dynamic point culling, the culling process was further optimized, including:

[0147] (4.1) Time Series Consistency Check: Detect the consistency of dynamic point changes in consecutive frames. If a feature point is frequently marked as a dynamic point in multiple consecutive frames, the culling strategy for this area is strengthened:

[0148] To improve the stability of dynamic point culling, time consistency checking was used to verify the dynamic feature points in consecutive frames. The specific algorithm is as follows:

[0149] First, define the dynamic state sequence of feature points in consecutive frames: ; where represents that the feature point is marked as a dynamic point in the t-th frame, represents a static point.

[0150] Second, for each feature point, calculate its time consistency score: ; where T is the size of the time window. If ( is the dynamic point threshold), then it is considered that is a dynamic point and is culled, otherwise it is retained as a static point.

[0151] Finally, perform a secondary verification on the feature points with abnormal time consistency check results, and determine whether to cull by statistically judging the consistency of the movement directions of point trajectories: ; ; if , then cull this point.

[0152] (4.2) Weighted Culling: Assign weights to feature point culling according to the confidence of the YOLOv11 detection boxes. Feature points within high-confidence detection boxes are preferentially culled, while feature points in low-confidence regions are further judged.

[0153] The embodiments are as follows:

[0154] Taking a frame of dynamic environment image as an example, the dynamic point culling process of the present invention is as follows:

[0155] Input image , after preprocessing, it is sent to the YOLOv11 model to obtain the detection result ;

[0156] Use the ORB algorithm to extract the feature point set ;

[0157] According to the bounding box of YOLOv11, it will satisfy The feature points of the judgment condition are from the set Remove from the list to get a set of static feature points ;

[0158] Will As input, it is combined with the descriptor and sent to the ORB-SLAM3 system for pose estimation and map construction.

[0159] Repeat the above process to process subsequent frame images.

[0160] Through this implementation, the present invention can effectively remove dynamic feature points in a dynamic environment, reduce the mismatch rate, and improve the accuracy and robustness of the SLAM system. Experiments show that after adopting the dynamic point removal method of the present invention, the errors in map construction and pose estimation are significantly reduced, especially in scenes with more dynamic interference.

[0161] like Figure 4 As shown, the present invention also provides a SLAM system for dynamic environment dense point cloud based on YOLOv11 model and ORB-SLAM3, and the system includes the following modules:

[0162] Dynamic object detection and segmentation module: The structured light system captures images in a dynamic environment and

[0163] The advanced YOLOv11 model is used for dynamic target detection and image segmentation;

[0164] Determine and remove dynamic feature points module: According to the detection results, determine whether the dynamic feature points are located in the dynamic

[0165] If the object is within the detection frame of a dynamic object, the dynamic feature points will be removed;

[0166] Pose estimation and map construction module: Input the remaining static feature points into the SLAM system for

[0167] Pose estimation and map construction; and evaluate and optimize the system robustness based on the pose estimation results;

[0168] Weighted elimination of dynamic feature points module: Use temporal consistency check to verify the dynamic feature points of consecutive frames and further weight them to complete pose estimation and mapping.

[0169] The present invention also provides an electronic device, comprising:

[0170] one or more processors;

[0171] A memory for storing one or more programs;

[0172] When the one or more programs are executed by the one or more processors, the one or more processors implement the SLAM method for dense point clouds in a dynamic environment based on the YOLOv11 model and ORB-SLAM3 as described above.

[0173] And a computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the SLAM method for dense point clouds in a dynamic environment based on the YOLOv11 model and ORB-SLAM3 as described above are implemented.

[0174] The above are only specific embodiments of the present application, enabling those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A SLAM method for dynamic environment dense point cloud based on YOLOv11 model and ORB-SLAM3, characterized in that: The method comprises the following steps: (1) Capturing images with a structured light system in a dynamic environment, and performing dynamic target detection and image segmentation based on an improved YOLOv11 model; the improved YOLOv11 model includes an input layer, a preprocessing layer, a plurality of standard convolutional layers, a GSConv layer, a VoVGSCSP feature fusion layer, and an output layer, wherein the GSConv layer is used to replace a traditional convolutional layer to balance accuracy and computational load, and the VoVGSCSP feature fusion layer is used to replace a C2f feature fusion layer; (2) According to the detection result, determine whether the dynamic feature point is located in the detection frame of the dynamic object. If it is located in the detection frame of the dynamic object, remove the dynamic feature point; (3) Input the remaining static feature points into the SLAM system for pose estimation and map construction; and evaluate and optimize the system robustness based on the pose estimation results; (4) Using temporal consistency check to verify the dynamic feature points of consecutive frames and further weightedly eliminate them to complete pose estimation and mapping; specifically, the following steps are included: (4.1) Time series consistency check Due to the consistency of the changes in dynamic points detected in the previous and next frames; when a feature point is marked as a dynamic point in multiple consecutive frames, a strategy is implemented to eliminate the area, specifically: the dynamic feature points of consecutive frames are verified using a temporal consistency check, as follows: First, define the dynamic state sequence of feature points in consecutive frames: in Indicates the feature point p in the tth frame i are marked as dynamic points. Represented as a static point; Secondly, for each feature point, calculate its temporal consistency score: Where T is the time window size; when C i ≥τ c , τ c is the dynamic point threshold, then p i It is a dynamic point and is removed, otherwise it is retained as a static point; Finally, the feature points with abnormal time consistency check results are verified twice, and the consistency of the movement direction of the statistical point trajectory is used to determine whether to remove them: If V i ≥τ v , then remove the point; (4.2) Weighted elimination: The feature points are weighted according to the confidence of the YOLOv11 detection frame; the feature points in the high-confidence detection frame are eliminated first, while the feature points in the low-confidence area are further judged.

2. The method according to claim 1, characterized in that The step (1) comprises the following steps: (1.1) Multi-scale feature extraction Extract feature maps from different levels of input captured by the structured light system: in, H l ,W l ,C l are the height, width and number of channels of the feature map respectively; (1.2) Spatial pyramid pooling: The multi-scale feature extraction result graph of the input layer is aggregated through pooling operations of different scales. Output multi-scale features: pool ki (x) = max(x i,j ), x i,j For all x i,j In k i In a window of size; where pool ki (·) indicates that the kernel size is k i Pooling operation; The SPP output features are processed by the GSConv convolution layer. GSConv combines standard convolution and depth-wise separable convolution with a weight parameter of W. The output is: F GCn =s; ReLU activation function σ(x): σ(x)=max(0,x); The final fusion output features: F fso =concat; (1.3) Input layer The spatial pyramid pooling results are normalized and sent to the network for forward propagation; The input image is adjusted to the standard size required by the YOLOv11 model and normalized to fit the model input requirements; the normalization formula is: Among them, I raw is the original image pixel value, I norm is the normalized pixel value; (1.4) Object Detection The model is used to generate dynamic target detection results, including target category labels, confidence scores, and bounding boxes, where each bounding box is defined as Dynamic objects in The data flow of the model first convolves the multi-scale feature image to reduce the number of channels of the original C1 to C2, which is named T1. Then, T1 is subjected to a depth-wise separable convolution operation to extract the deep features and named T2; Finally, T1 and T2 are merged to form a result with twice the number of channels of C2, named T3; Finally, T3 is reshuffled according to the eigenvalues: FSPP=concat(poolk1(Ft),poolk2(Ft),…,poolkn(Ft)); The output of the dynamic target detection results generated by the model includes the following information: (a) Target category label c, indicating the category of the detected target; (b) Confidence s, indicating the probability that the target belongs to category c; (c) Bounding box Circumference, of which: x min ,y min : Coordinate of the upper left corner of the bounding box; x max ,y max : Coordinates of the lower right corner of the bounding box; Combining the results, we can express it as: Among them, i is the index of the target and N is the total number of targets detected in the current frame.

3. The method according to claim 1, characterized in that The step (2) specifically comprises the following steps: (2.1) Extract feature points and their descriptors from the current frame using the ORB algorithm. Feature points are represented by pixel coordinates (x, y). (2.2) For each dynamic target detected in step (1), obtain its bounding box coordinates and mark all feature points in the bounding box as dynamic feature points; the determination conditions of dynamic feature points are as follows: x min ≤x≤x max ,and min ≤y≤y max ; When (x, y) meets the above conditions, it is determined to be a dynamic point, otherwise it is retained as a static point; (2.3) For the set of static feature points after elimination, their coordinates and descriptors are retained for subsequent processing of the SLAM system.

4. The method according to claim 1, characterized in that The step (3) specifically comprises the following steps: (3.1) Pose estimation Preliminary estimation of camera pose based on static feature points; ORB-SLAM3 calculates camera pose through feature point matching and triangulation method: For a static feature point set: P static ={p1,p2,…,p n }; And the set of feature points in the keyframe: Q={q1,q2,…,q m }; To perform descriptor matching, the expression for calculating the similarity between two points using the Hamming distance is as follows: in, represents the XOR operation, popcount is a function that counts the number of 1s, and retains the smallest distance that is lower than the threshold τ d The matching pair (p,q); (3.2) Pose estimation between two frames Robust estimation is performed through the two models of basic matrix F and homography matrix H, and the RANSAC algorithm is used to filter the mismatched point pairs to obtain a set of matching points that meet the geometric constraints; the basic matrix F is used to describe the epipolar geometric relationship between the point pairs in the two images. For the matching point pairs (x i ,x i ′ ),satisfy: in and Represents the homogeneous coordinates of the same scene point in two images; The homography matrix H describes the perspective transformation between two images and is used for pure rotation or planar scenes. i ,x i ′ ),satisfy: x i ′ ~Hx i ; Right now: x i ′ =λHx i ; Where λ is a scaling factor; (3.3) Optimizing pose estimation According to the set of matching points, the pose T of the current frame is estimated through the PnP problem cw , whose optimization goal is to minimize the reprojection error: AND reproj =∑ i |p i -π(K·[R|t]·P i )| 2 ; Among them, p i =(x i ,y i ) is the feature point in the image coordinate system; P i =(X i ,Y i ,Z i ,1) T is the feature point in three-dimensional space, K is the camera intrinsic parameter matrix, [R|t] is the camera extrinsic parameter, and π is the projection function: π([u,v,w] T )=(u / w,v / w); Solve E by Gauss-Newton method or LM algorithm reproj The minimum value of cw ; (3.4) Optimize map and posture Construct a model based on graph optimization, optimize the map and pose by minimizing the error cost function, and eliminate abnormal errors caused by residual dynamic points. In global optimization, use a graph-based optimization method to jointly optimize the pose and map, with the goal of minimizing the error cost function of the entire graph model: Among them, e i is the observation error, z i is the observed value, h(x i ) is an estimated value, is the observation error covariance matrix; Solve the above cost function through the G2O tool or Ceres Solver to optimize the node pose and map point position.

5. The method according to claim 4, characterized in that The step (3.2) also includes using the RANSAC algorithm to perform model fitting by randomly sampling point pairs to achieve robust model estimation. The specific steps are as follows: (a) Randomly extract the minimum point set from the matching point set Used to calculate the candidate fundamental matrix FFF or homography matrix HHH; (b) Use the fitted model to evaluate all matching points {(x i ,x i ′ )}Geometric error: For the fundamental matrix: Satisfy F (x i ,x i ′ )<∈ is an internal point; where ∈ is a manually input threshold; For the homography matrix: d H (x i ,x i ′ )=|x i ′ -λHx i |2; Satisfy H (x i ,x i ′ )<∈ is an interior point. (c) Calculate the number of inliers and select the model with the most inliers as the final estimation result; Through the above steps, the mismatched point pairs are filtered out and the point pairs that meet the geometric constraints are retained: H (x i ,x i ′ )<∈.

6. A SLAM system for dynamic environment dense point cloud based on YOLOv11 model and ORB-SLAM3, characterized in that: The system includes the following modules: Dynamic target detection and segmentation module: The structured light system captures images in a dynamic environment, and performs dynamic target detection and image segmentation based on the improved YOLOv11 model; the improved YOLOv11 model includes an input layer, a preprocessing layer, multiple standard convolutional layers, a GSConv layer, a VoVGSCSP feature fusion layer, and an output layer. The GSConv layer is used to replace the traditional convolutional layer to balance accuracy and computational load, and the VoVGSCSP feature fusion layer is used to replace the C2f feature fusion layer. Determine and remove dynamic feature points module: According to the detection results, determine whether the dynamic feature points are located in the detection frame of the dynamic object. If they are located in the detection frame of the dynamic object, remove the dynamic feature points; Pose estimation and map construction module: input the remaining static feature points into the SLAM system for pose estimation and map construction; and evaluate and optimize the system robustness based on the pose estimation results; Weighted dynamic feature point elimination module: Use temporal consistency check to verify the dynamic feature points of consecutive frames and further weighted eliminate them to complete pose estimation and map construction; specifically: (a) Time series consistency check Due to the consistency of the changes in dynamic points detected in the previous and next frames; when a feature point is marked as a dynamic point in multiple consecutive frames, a strategy is implemented to eliminate the area, specifically: the dynamic feature points of consecutive frames are verified using a temporal consistency check, as follows: First, define the dynamic state sequence of feature points in consecutive frames: in Indicates the feature point p in the tth frame i are marked as dynamic points. Represented as a static point; Secondly, for each feature point, calculate its temporal consistency score: Where T is the time window size; when C i ≥τ c , τ c is the dynamic point threshold, then p i It is a dynamic point and is removed, otherwise it is retained as a static point; Finally, the feature points with abnormal time consistency check results are verified twice, and the consistency of the movement direction of the statistical point trajectory is used to determine whether to remove them: If V i ≥τ v , then remove the point; (b) Weighted elimination: The feature points are weighted according to the confidence of the YOLOv11 detection frame; the feature points in the high-confidence detection frame are eliminated first, while the feature points in the low-confidence area are further judged.

7. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the SLAM method of dynamic environment dense point cloud based on the YOLOv11 model and ORB-SLAM3 as described in any one of claims 1-5.

8. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by the processor, the steps of the SLAM method of dynamic environment dense point cloud based on the YOLOv11 model and ORB-SLAM3 as described in any one of claims 1 to 5 are implemented.

Images