ATY-SLAM: A Visual Semantic SLAM for Dynamic Indoor Environments

Visual Simultaneous Localization and Mapping (VSLAM) is a critical technology that enables mobile robots to accurately sense their surroundings and perform localization and map building. However, the assumptions underlying VSLAM algorithms are based on static environments, often leading to poor performance in highly dynamic indoor scenes. Accurately estimating camera pose and achieving precise localization in such environments poses a significant challenge. This paper presents ATY-SLAM: Adaptive Thresholding combining YOLOv7-tiny SLAM, a VSLAM method for dynamic feature point culling and keyframe optimization in highly dynamic scenes. It can effectively improve robustness and accuracy in dynamic indoor environments. In ATY-SLAM system, we employ a combination of the YOLOv7-tiny object detection network, motion consistency detection, and the LK optical flow algorithm to detect dynamic regions in the image. Then, the unstable feature points are removed, thus further improving the stability of the VSLAM system. We then employ an adaptive thresholding method to select a stable keyframe, which solves the problem of poor quality of keyframes selected by existing heuristic thresholding. Experimental results on the public TUM RGB-D dataset demonstrate that the proposed algorithm reduces the absolute trajectory error in high dynamic scenes by an average of 96.4% compared to ORB-SLAM3. Additionally, while maintaining similar accuracy, the time required to process each frame in the tracking thread is reduced by over 98.8% compared to the classical DynaSLAM algorithm, achieving real-time performance.