Dynamic environment dense point cloud SLAM method and system based on YOLOv11 and ORB-SLAM3

Abstract

The invention discloses a dynamic environment dense point cloud SLAM (Simultaneous Localization and Mapping) method and system based on YOLOv11 and ORB-SLAM 3. According to the method, real-time target detection and image segmentation technologies of a YOLOv11 model are integrated into an ORB-SLAM 3 framework, so that high-precision and robust visual SLAM in a dynamic environment is realized. A traditional convolution layer is replaced by using a balanced convolution method GSConv layer in a YOLOv11 model, and a traditional C2f layer is replaced by using a new feature fusion module VoVGSCS layer, so that a Neck structure of the YOLOv11 is improved, and a lightweight network model is realized. Experimental data prove that the pose estimation precision of the method in a dynamic environment is remarkably superior to that of an existing visual SLAM algorithm.

Description

Technical Field

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

[0002] With the development of society and the advancement 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 dynamic environments remains a major challenge. This paper aims to propose a method based on improved YOLOv11 fused with ORB-SLAM3 to cope with the dense point cloud SLAM problem in dynamic environments. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a SLAM method and system for dynamic environment dense point cloud based on YOLOv11 and ORB-SLAM3, the system includes three modules: feature extraction, feature matching and feature point removal. The technical solution of the invention is achieved by the following contents: The first aspect of the present invention: a SLAM method for dynamic environment dense point cloud based on YOLOv11 model and ORB-SLAM3, the method comprising the following steps: (1) The structured light system captures images in a dynamic environment and performs dynamic target detection and image segmentation based on the improved YOLOv11 model; (2) According to the detection results, 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, the dynamic feature point is removed; (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) The dynamic feature points of consecutive frames are verified by temporal consistency check and further weighted elimination to complete pose estimation and mapping.

[0004] 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.

[0005] Furthermore, the step (1) includes the following steps: (1.1) Multi-scale feature extraction Extract feature maps from different levels of input captured by the structured light system: ; in, , 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: , For all exist In a window of size The kernel size is Pooling operation; The SPP output features are processed by the GSConv convolution layer. GSConv combines standard convolution and depth-wise separable convolution with weight parameters of , the output is: ; Activation Function : ; The final fusion output features: ; (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: ;in, is the original image pixel value, 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 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; 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: 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: : 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.

[0006] Furthermore, the step (2) specifically includes the following steps: (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; (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; (2.3) For the set of static feature points after elimination, retain their coordinates and descriptors for subsequent processing of the SLAM system.

[0007] Furthermore, the step (3) specifically includes 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: ; And the set of feature points in the keyframe: ; To perform descriptor matching, the expression for calculating the similarity between two points using the Hamming distance is as follows: ;in, represents the exclusive OR operation, It is a function that counts the number of 1s, retaining the smallest distance that is below the threshold Matching pairs ; (3.2) Pose estimation between two frames Robust estimation is performed through two models: the basic matrix F and the homography matrix H. The RANSAC algorithm is used to filter the mismatched point pairs to obtain a set of matching points that meet the geometric constraints. Used to describe the epipolar geometric relationship between point pairs in two images. ,satisfy: ;in and Represents the homogeneous coordinates of the same scene point in two images; The homography matrix Describes the perspective transformation between two images, which is used for pure rotation or planar scenes; for matching point pairs ,satisfy: ;Right now: ;in is a scaling factor; (3.3) Optimizing pose estimation Based on the set of matching points, the pose of the current frame is estimated through the PnP problem , whose optimization goal is to minimize the reprojection error: ; in, is the feature point in the image coordinate system; is the feature point in three-dimensional space, K is the camera intrinsic parameter matrix, is the camera extrinsic parameter, is the projection function: ; Solved by Gauss-Newton method or LM algorithm The minimum value of ; (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: ;in, is the observation error, is the observed value, is an estimate, 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.

[0008] Furthermore, the step (3.2) further includes adopting the RANSAC algorithm to implement robust model estimation by performing model fitting through random sampling point pairs. 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 Or the homography matrix ; (b) Evaluate all matching points using the fitted model The geometric error is: For the fundamental matrix: ; satisfy The point pairs of are interior points; To manually enter the threshold; For the homography matrix: ; satisfy The point pairs of are interior points; (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 set of point pairs that meet the geometric constraints is retained: .

[0009] Furthermore, the step (4) includes the following steps: (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 time consistency check, as follows: First, define the dynamic state sequence of feature points in consecutive frames: ; in Indicates the feature point in the tth frame 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 , is the dynamic point threshold, then it is considered 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: ; ;like , then remove the point; (4.2) Weighted elimination: Weights are assigned to feature point elimination according to the confidence of the YOLOv11 detection frame; feature points in high-confidence detection frames are eliminated first, while feature points in low-confidence areas are further judged.

[0010] The second aspect of the present invention: a SLAM system for dynamic environment dense point cloud based on YOLOv11 model and ORB-SLAM3, the system comprising 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; 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 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.

[0011] A third aspect of the present invention: an electronic device, comprising: 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 YOLOv11 model and ORB-SLAM3.

[0012] The fourth aspect of the present invention: a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the SLAM method for dynamic environment dense point cloud based on the YOLOv11 model and ORB-SLAM3.

[0013] The beneficial effects of the present invention are as follows: This paper uses the high-precision target detection capability of the YOLOv11 model to remove dynamic feature points from the scene, thereby effectively reducing the interference of dynamic objects on the visual SLAM system. By introducing this removal method in the visual odometry stage of ORB-SLAM3, the mismatching problem caused by moving objects in dynamic scenes is solved, and the positioning accuracy and trajectory stability of the system are effectively improved. In order to better meet the needs of real-time SLAM systems, this paper makes targeted improvements to YOLOv11. Specifically, the GSConv convolution technology and the VoVGSCSP feature fusion layer are introduced, which greatly improves the lightweight level and inference speed of the model, while maintaining a high detection accuracy in dynamic target detection tasks. Experimental results show that the improved YOLOv11 effectively reduces the computational complexity of the system while maintaining detection performance, laying the foundation for real-time dynamic point removal. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The attached figures are incorporated into the specification and serve as an integral part of the specification, illustrating the implementation methods of the various modules involved in the present invention in accordance with the principles of the present invention. The focus of the drawings is not to limit but to explain the principles of the invention. In the drawings, Figure 1 It is an execution flow chart of each module of the dynamic feature point removal system of the present invention; Figure 2 It is a data flow diagram of the GSConv convolutional layer of the present invention; Figure 3 is an epipolar geometry diagram in the process of estimating the pose between two frames of the present invention; Figure 4 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0015] The following detailed description refers to the drawings and will set forth specific details of an embodiment in order to provide a comprehensive understanding of the various aspects of the invention for which protection is sought. For developers who are familiar with the relevant technology, it is obvious that other embodiments different from the specific details described below can be used to implement certain modules and systems in the present invention. The following description is intended to be illustrative rather than limiting. All modifications, equivalent substitutions, etc. made within the spirit and principles of the invention should be included in the scope of protection of the invention.

[0016] According to the solution of the present invention, the designed improved YOLOv11 and fusion ORB-SLAM3 system modules and processes are as follows Figure 1 As shown. The present invention applies the scheme to a structured light system. The specific scheme is as follows: S1: Input image preprocessing and object detection (1) Image preprocessing (1.1) Multi-scale feature extraction Extract feature maps from different levels of input captured by the structured light system: ;in, , are the height, width and number of channels of the feature map respectively.

[0017] (1.2) Spatial pyramid pooling: The multi-scale feature extraction result graph of the input layer is aggregated through pooling operations of different scales to output multi-scale features , For all exist In a window of size The kernel size is The pooling operation.

[0018] The SPP output features are processed by the GSConv convolution layer. GSConv combines standard convolution and depth-wise separable convolution with weight parameters of , the output is: ; Activation Function : ; Final fusion output features: ; (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: ;in, is the original image pixel value, is the normalized pixel value.

[0019] (1.4) Object Detection The model is used to generate dynamic target detection results, including target category labels, confidence levels, and bounding boxes, where each bounding box is defined as Dynamic objects in .

[0020] The data flow of the model is as follows Figure 2 As shown, first, the multi-scale feature image is convolved to reduce the number of channels of the original C1 to C2, and 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.

[0021] (b) Confidence s, which indicates the probability that the target belongs to category c.

[0022] (c) Bounding box Circumference, of which: : Coordinates of the upper left corner of the bounding box.

[0023] : Coordinates of the lower right corner of the bounding box.

[0024] Combining these results, it can be expressed as: ; Where: i is the index of the target, and N is the total number of targets detected in the current frame.

[0025] S2: Identification and elimination of dynamic feature points According to the target detection results of step S1, the dynamic area in the image is matched with the feature points extracted by ORB-SLAM3 one by one, and the feature points falling in the dynamic area are removed. The specific steps are as follows: (2.1) Use the ORB algorithm to extract feature points and their descriptors from the current frame. Feature points are represented by pixel coordinates. express.

[0026] (2.2) For each dynamic target detected by S1, obtain its bounding box coordinates and mark all feature points in 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 to be a dynamic point, otherwise it is retained as a static point.

[0027] (2.3) For the set of static feature points after elimination, retain their coordinates and descriptors for subsequent processing of the SLAM system.

[0028] S3: Input of static feature points and SLAM processing Input the static feature points after removing the dynamic points into ORB-SLAM3 and complete the following operations: (3.1) Pose Estimation The camera pose is preliminarily estimated based on static feature points. ORB-SLAM3 calculates the camera pose through feature point matching and triangulation method: For the static feature point set: ; and the set of feature points in the keyframe: ; Perform descriptor matching and use the Hamming distance to calculate the similarity between two points: ;in, represents the exclusive OR operation, It is a function that counts the number of 1s. The minimum distance is retained and is below the threshold. Matching pairs .

[0029] (3.2) Pose estimation between two frames Robust estimation is performed using the two models of the basic matrix F and the homography matrix H, and the RANSAC algorithm is used to filter out mismatched point pairs to obtain a set of matching points that meets the geometric constraints: Basic Matrix It is used to describe the epipolar geometric relationship between point pairs in two images. The epipolar relationship is as follows Figure 3 As shown. For the matching point pair ,satisfy: in and Represents the homogeneous coordinates of the same scene point in two images.

[0030] Homography matrix Describes the perspective transformation between two images and is applicable to pure rotation or planar scenes. ,satisfy: ; Right now: ;in is a scaling factor.

[0031] In order to achieve robust model estimation, the RANSAC algorithm is used to fit the model by randomly sampling point pairs. The specific steps are as follows: Randomly extract the minimum point set from the matching point set , used to calculate the candidate fundamental matrix F F or homography matrix H H.

[0032] Evaluate all matching points using the fitted model The geometric error is: For the fundamental matrix:

[0033] satisfy The point pairs of are interior points.

[0034] For the homography matrix ; satisfy The point pairs of are interior points.

[0035] Calculate the number of inliers and select the model with the most inliers as the final estimation result.

[0036] Through the above steps, the mismatched point pairs are filtered out and the set of point pairs that meet the geometric constraints is retained: ; in The threshold is manually input, which is determined according to the accuracy of the structured light and the stability of the inertial navigation dynamic structure.

[0037] (3.3) Optimizing pose estimation Based on the set of matching points, the pose of the current frame is estimated through the PnP (Perspective-n-Point) problem , whose optimization goal is to minimize the reprojection error: ; in, is the feature point in the image coordinate system; is the feature point in three-dimensional space, K is the camera intrinsic parameter matrix, is the camera extrinsic parameter, is the projection function: .

[0038] Solved by Gauss-Newton method or LM (Levenberg-Marquardt) algorithm The minimum value of .

[0039] (3.4) Optimize map and posture Build a graph optimization-based model to optimize the map and pose by minimizing the error cost function, and eliminate abnormal errors that may be caused by residual dynamic points. In global optimization, a graph-based optimization method is used to jointly optimize the pose and map, with the goal of minimizing the error cost function of the entire graph model: ;in, is the observation error, is the observed value, is an estimate, is the observation error covariance matrix.

[0040] Solve the above cost function through the G2O (General Graph Optimization) tool or Ceres Solver to optimize the node pose and map point position.

[0041] S4: Performance optimization and improved strategies for dynamic point culling To improve the accuracy of dynamic point culling, the culling process was further optimized, including: (4.1) Time series consistency check: Check the consistency of dynamic point changes in the previous and next frames. If a feature point is frequently marked as a dynamic point in multiple consecutive frames, strengthen the elimination strategy for this area: In order to improve the stability of dynamic point removal, the dynamic feature points of consecutive frames are verified by using temporal consistency check. The specific algorithm is as follows: First, define the dynamic state sequence of feature points in consecutive frames: ;in Indicates the feature point in the tth frame are marked as dynamic points. Represented as a static point.

[0042] Secondly, for each feature point, calculate its temporal consistency score: ; Where T is the time window size. If ( is the dynamic point threshold), then it is considered If it is a dynamic point, it will be removed; otherwise, it will be kept as a static point.

[0043] 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: ; ;like , then remove the point.

[0044] (4.2) Weighted elimination: Weights are assigned to feature point elimination based on the confidence of the YOLOv11 detection frame. Feature points in high-confidence detection frames are eliminated first, while feature points in low-confidence areas are further judged.

[0045] The examples are as follows: Taking a frame of dynamic environment image as an example, the dynamic point removal process of the present invention is as follows: Input Image , after preprocessing, it is sent to the YOLOv11 model to obtain the detection results ; Use ORB algorithm to extract feature point set ; 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 ; Will As input, it is combined with the descriptor and sent to the ORB-SLAM3 system for pose estimation and map construction.

[0046] Repeat the above process to process subsequent frame images.

[0047] 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.

[0048] like Figure 4 As shown, the present invention also provides a SLAM system of a dynamic environment dense point cloud based on the YOLOv11 model and ORB-SLAM3, the system comprising the following modules: Dynamic object detection and segmentation module: The structured light system captures images in a dynamic environment and The advanced YOLOv11 model is used for dynamic target detection and image segmentation; Determine and remove dynamic feature points module: According to the detection results, determine whether the dynamic feature points are located in the dynamic If the object is within the detection frame of a dynamic object, the dynamic feature points will be removed; 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 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.

[0049] The present invention also provides an electronic device, comprising: 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 YOLOv11 model and ORB-SLAM3 as described.

[0050] And a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the SLAM method for dynamic environment dense point cloud based on the YOLOv11 model and ORB-SLAM3.

[0051] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent 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 the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features applied for 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) The structured light system captures images in a dynamic environment and performs dynamic target detection and image segmentation based on the improved YOLOv11 model; (2) According to the detection results, 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, the dynamic feature point is removed; (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) The dynamic feature points of consecutive frames are verified by temporal consistency check and further weighted elimination to complete pose estimation and mapping.

2. The method according to claim 1, characterized in that 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.

3. 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, , 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: , For all exist In a window of size The kernel size is Pooling operation; The SPP output features are processed by the GSConv convolution layer. GSConv combines standard convolution and depth-separable convolution with weight parameters of , the output is: ; Activation Function : ; The final fusion output features: ; (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: ;in, is the original image pixel value, 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 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; 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: 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: : 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.

4. The method according to claim 1, characterized in that: The step (2) specifically includes the following steps: (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; (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; (2.3) For the set of static feature points after elimination, retain their coordinates and descriptors for subsequent processing of the SLAM system.

5. The method according to claim 1, characterized in that The step (3) specifically includes 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: ; And the set of feature points in the keyframe: ; To perform descriptor matching, the expression for calculating the similarity between two points using the Hamming distance is as follows: ;in, Represents the exclusive OR operation, It is a function that counts the number of 1s, retaining the smallest distance that is below the threshold Matching pairs ; (3.2) Pose estimation between two frames Robust estimation is performed through two models: the basic matrix F and the homography matrix H. The RANSAC algorithm is used to filter the mismatched point pairs to obtain a set of matching points that meet the geometric constraints. Used to describe the epipolar geometric relationship between point pairs in two images. ,satisfy: ;in and Represents the homogeneous coordinates of the same scene point in two images; The homography matrix Describes the perspective transformation between two images, which is used for pure rotation or planar scenes; for matching point pairs ,satisfy: ;Right now: ;in is a scaling factor; (3.3) Optimizing pose estimation Based on the set of matching points, the pose of the current frame is estimated through the PnP problem , whose optimization goal is to minimize the reprojection error: ; in, is the feature point in the image coordinate system; is the feature point in three-dimensional space, K is the camera internal parameter matrix, is the camera extrinsic parameter, is the projection function: ; Solved by Gauss-Newton method or LM algorithm The minimum value of ; (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: ;in, is the observation error, is the observed value, is an estimate, 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.

6. The method according to claim 5, 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 Or the homography matrix ; (b) Evaluate all matching points using the fitted model The geometric error is: For the fundamental matrix: ; satisfy The point pairs of are interior points; To manually enter the threshold; For the homography matrix: ; satisfy The point pairs of are interior points; (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 pair set that meets the geometric constraints is retained: .

7. The method according to claim 1, characterized in that The step (4) comprises the following steps: (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 in the tth frame 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 , is the dynamic point threshold, then it is considered 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: ; ;like , then remove the point; (4.2) Weighted elimination: Weights are assigned to feature point elimination according to the confidence of the YOLOv11 detection frame; feature points in high-confidence detection frames are eliminated first, while feature points in low-confidence areas are further judged.

8. 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; 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 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.

9. 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 1-7.

10. 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 are implemented as described in any one of claims 1-7.

Images