DOTF-SLAM: Real-Time Dynamic SLAM Using Dynamic Odject Tracking and Key-Point Filtering

Traditional visual simultaneous localization and mapping (SLAM) algorithms assume static scenes, which limits their application in real-world environments where dynamics are prevalent, such as autonomous driving and multi-robot collaboration. Therefore, clear information about the dynamic environment is needed to aid decision-making and scene understanding. To address the problem, this paper develops a method based on the ORB-SLAM2 framework that is more robust when operating in dynamic environments. In our method, we combine dynamic object tracking, prediction and dynamic feature points filtering to eliminate the influence of dynamic objects on localization and map construction. On the TUM dataset, the algorithm reduces the Absolute Trajectory Error (ATE) by more than 80% compared to ORB-SLAM2, while the improvement in dynamic segments of the KITTI dataset is around 20%. In addition, we achieve a real-time performance of over 15 FPS while localization accuracy is comparable to DynaSLAM and DS-SLAM, which can only achieve approximately 2-3 FPS. According to the experimental results, suggested algorithm can successfully improve localization accuracy in highly dynamic situations.