DynPL-SLAM: A Robust Stereo Visual SLAM System for Dynamic Scenes Using Points and Lines

Abstract

Simultaneous Localization and Mapping (SLAM) plays a crucial role in enabling intelligent mobile robots and vehicles to estimate their state in unknown environments. However, most visual SLAM systems are built on the assumption of static scenes, which limits their performance in dynamic environments. This paper proposes DynPL-SLAM, an indirect stereo visual SLAM system that leverages both point and line features to handle dynamic scenes. Our approach introduces novel dynamic feature detectors to effectively deals with dynamic objects in highly dynamic scenes. Furthermore, we propose a new representation of spatial line residual error and provide a comprehensive explanation of its integration into pose optimization. Additionally, we propose a novel Histogram of Regional Similarity (HRS) model which fast computes the similarity between scenes, and it is used in keyframe selection and online loop closure detection (LCD). In order to evaluate the performance of DynPL-SLAM, we conduct comparative analyses against existing state-of-the-art (SOTA) approaches using various publicly available datasets and our own recorded datasets. The experimental results demonstrate that our approach significantly improves localization accuracy in dynamic scenes and outperforms existing stereo visual SLAM systems in terms of real-time performance. In fact, our approach achieves approximately 18.6% faster processing speed than ORB-SLAM3.