Visual SLAM combining YOLO11n-seg and bidirectional LK optical flow for dynamic environments

Abstract

To improve the localization accuracy and mapping quality of visual SLAM in highly dynamic scenes, this paper proposes a visual SLAM algorithm (YLK-SLAM) fusing YOLO11n-seg and the LK bidirectional optical flow method. Based on the ORB-SLAM3 framework, the algorithm employs the lightweight YOLO11n-seg for feature extraction and semantic perception at the front end, and uses an improved bidirectional optical flow tracking algorithm to eliminate dynamic objects in a timely manner at the back end. Comparative experiments are conducted on the TUM and EuRoC datasets. The proposed YLK-SLAM algorithm can effectively eliminate dynamic features. Especially in high-dynamic scenarios, compared with traditional SLAM algorithms (ORB-SLAM2, VDO-SLAM), it shows significant improvements in positioning accuracy and running time, effectively suppressing the adverse impact of dynamic features on positioning and mapping accuracy, and constructing a globally consistent map.