Robust Tracking and Localization using Stable Features of Static Planes in Dynamic Scenes

Dynamic interference by moving objects and localization failure induced by low-texture scenes are two urgent problems to be solved in the practical applications of visual simultaneous localization and mapping (SLAM). This study proposes a robust tracking and localization approach with a monocular camera sensor that leverages points within extracted lines and planes to deal with the challenges of dynamic objects and low-texture features. Firstly, we use a line segment detector and an improved CNN-based plane segmentation network to extract lines and planes, respectively, without detecting dynamic objects, thus avoiding complex semantic models. Then, the tracking thread tightly uses plane and line information to promote front-end pose tracking. Finally, the geometric constraint is added to the BA constraints in the pose optimization stage to enhance the robustness. We have evaluated the proposed approach on public datasets and real-scene experiments. Results demonstrate that our method outperforms several state-of-the-art dynamic filtering-based methods in highly dynamic scenes, significantly improving robustness and positioning accuracy.