FD-SLAM: A Semantic SLAM Based on Enhanced Fast-SCNN Dynamic Region Detection and DeepFillv2-Driven Background Inpainting

Abstract

Semantic SLAM integrates semantic networks into SLAM systems to ensure the proper functioning of mobile robots by detecting and removing dynamic areas in the environment. Existing semantic SLAM approaches face challenges such as map information loss and tracking failures due to the abandonment of background information obscured by dynamic objects. In this study, we propose an FD-SLAM system, which builds upon ORB-SLAM3 and combines improved Fast-SCNN and Deepfillv2. This system employs semantic segmentation and geometric methods to identify and eliminate dynamic regions while incorporating an efficient inpainting network, utilizing free-form masks, to fill in the background occluded by dynamic objects. This complete background information augmentation enables the operation of the SLAM system in dynamic environments. The evaluation on the public TUM dataset demonstrates that FD-SLAM outperforms other dynamic SLAM methods by maintaining tracking accuracy while achieving a higher success rate in tracking. This indicates that FD-SLAM can enhance the robustness of the system in dynamic environments.