Dynamic visual SLAM algorithm for urban forest environments based on semantic segmentation

Abstract

Abstract Simultaneous localization and mapping (SLAM) is a key technology for mobile robots operating in unknown environments. Most existing SLAM technologies assume that unknown scenarios are static, which are effective in indoor environments, but face limitations when a dynamic object exists. This challenge is particular evident in urban forest understory environments, where moving objects such as pedestrian significantly impact SLAM performance. To solve this problem, a dynamic visual SLAM algorithm for urban forests, referred to as YSLK-SLAM was proposed in this study. The algorithm was based on ORB-SLAM2, and a parallel semantic segmentation thread was integrated and the YOLOv8s-seg semantic segmentation model was used to infer RGB images, obtaining the mask of dynamic objects. At the same time, a missed detection module and an overlapping depth difference strategy were introduced to achieve refined segmentation of dynamic objects. By fusing the masks with the images, interference feature points were effectively mitigated. Additionally, the original tracking thread of ORB-SLAM2 was optimized with an optical flow tracking strategy for non-keyframes, enabling real-time operation in understory environments. Finally, the performance of YSLK-SLAM was evaluated against other SLAM algorithms using both the TUM dataset and a self-constructed urban forest dataset. Experimental results show that YSLK-SLAM achieved the lowest absolute trajectory error (ATE) across all test sequences in the TUM dataset. Compared to ORB-SLAM2 and ORB-SLAM3, the average error was reduced by 75.59% and 68.72%, respectively, with even more significant improvements in highly dynamic environments, where the error reduction reached 96.93% and 94.74%, respectively. Furthermore, in the urban forest dataset, YSLK-SLAM achieved the lowest average ATE of approximately 0.108 m across three test sequences, significantly outperforming ORB-SLAM2, ORB-SLAM3, Dyna-SLAM, and DS-SLAM. Additionally, the YSLK-SLAM maintained a frame rate of 30.3 FPS, ensuring high real-time performance. The proposed YSLK-SLAM provides a reliable solution for autonomous navigation of mobile robots or unmanned aerial vehicle in urban forest environments.