DFD-SLAM: Visual SLAM with Deep Features in Dynamic Environment

Abstract

Visual SLAM technology is one of the important technologies for mobile robots. Existing feature-based visual SLAM techniques suffer from tracking and loop closure performance degradation in complex environments. We propose the DFD-SLAM system to ensure outstanding accuracy and robustness across diverse environments. Initially, building on the ORB-SLAM3 system, we replace the original feature extraction component with the HFNet network and introduce a frame rotation estimation method. This method determines the rotation angles between consecutive frames to select superior local descriptors. Furthermore, we utilize CNN-extracted global descriptors to replace the bag-of-words approach. Subsequently, we develop a precise removal strategy, combining semantic information from YOLOv8 to accurately eliminate dynamic feature points. In the TUM-VI dataset, DFD-SLAM shows an improvement over ORB-SLAM3 of 29.24% in the corridor sequences, 40.07% in the magistrale sequences, 28.75% in the room sequences, and 35.26% in the slides sequences. In the TUM-RGBD dataset, DFD-SLAM demonstrates a 91.57% improvement over ORB-SLAM3 in highly dynamic scenarios. This demonstrates the effectiveness of our approach.