A Robust Visual SLAM System in Dynamic Environment

Simultaneous Localization and Mapping (SLAM) is vital for the navigation of autonomous robots in unknown environments. Current SLAM systems have made progress but still struggle to balance accuracy, robustness, and real-time processing in dynamic environments. This paper presents a visual SLAM system that, with the assistance of neural networks, significantly enhances localization accuracy in dynamic environments without compromising real-time performance. It utilizes a static feature point algorithm based on semantic information during feature extraction to isolate static from dynamic feature points for better tracking and integrates a fast feature point weight calculation algorithm that assesses feature reliability based on proximity to dynamic objects, which greatly enhances the initial camera pose. Moreover, a match detection algorithm removes incorrect match relationships during local map tracking, which boosts pose precision and system robustness. The experiments on the TUM datasets show our system's superior performance. Specifically, in the s/static data sequence, our system achieves over 51.9% improvement on the RPE RMSE metric, while other systems either do worse than ORB-SLAM3 or improve less than 26.4%. These results prove the robustness of our system.