Visual SLAM algorithm based on dynamic attribute recognition in dynamic scenes

Visual SLAM algorithms suffer from errors in dynamic scenes due to interference from dynamic objects, leading to significant degradation in system accuracy. Existing deep learning-based methods for removing feature points of dynamic objects focus solely on actively moving objects with prior information, neglecting the dynamic characteristics of passive objects. To address this issue, we propose the concept of dynamic attribute recognition based on active–passive relationships; starting from the correlation between objects, we determine the dynamic attributes of passive objects based on the dynamic attributes of active objects. Building upon this, we propose a visual SLAM algorithm based on dynamic attribute recognition. First, YOLO is employed for object detection to identify the semantic information of objects and obtain their positional information. Then, the Euclidean distance is used to determine the affiliation between active and passive objects. Specifically, if the active object is a person, the Euclidean distance between the passive object and the hand keypoints determines their affiliation. Simultaneously, the Lucas-Kanade optical flow method and RANSAC are used to further assist in determining the dynamic attributes of both active and passive objects. Finally, feature points within the regions occupied by active and passive objects identified as dynamic are either removed or their weights are reduced, relying on static feature points to achieve camera pose estimation. Experimental results demonstrate that our algorithm reduces both the Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) by over 90% compared to the original ORB-SLAM2 on high-dynamic sequences of the TUM dataset. Compared to similar algorithms such as DS-SLAM and Dyna-SLAM, our method exhibits superior accuracy and robustness in dynamic scenes containing both active and passive objects, with the tracking thread processing each frame in an average of only 29.36 ms. Our approach significantly enhances both the accuracy and real-time performance of SLAM algorithms in dynamic scenes.