R4-SLAM: Toward Real-time, Robust, and Resource-Restricted Visual SLAM in Dynamic Environments

Recent efforts have attempted to incorporate semantic information from neural networks into SLAM to mitigate the distraction from dynamic objects. However, due to the expensive computational costs of dynamic features, it is impractical to draw real-time implementation on resource-limited edge devices. This paper introduces a low-cost method, R4-SLAM, that appropriately balances the stability, speed, and accuracy of visual SLAM in dynamic environments. To further boost the speed of R4-SLAM towards real-time, we present a simple yet efficient semantic keyframe selection strategy that leverages redundant frames to preserve both stability and accuracy. Moreover, we introduce a semantic front-end decoupling framework to offset prolonged and time-consuming inference modules. Through comprehensive experiments, we demonstrate the advantages of R4-SLAM: achieving significantly higher processing speeds (14.71 FPS) and stability (99.16%) compared to the state-of-the-art SG-SLAM [1] (8.24 FPS and 79.86%) and SCP-SLAM [2] (5.17 FPS and 98.08%) while maintaining comparable accuracy on widely used TUM RGB-D datasets.