FDR-SLAM: A Visual SLAM with Fast Object Detection, Depth-Enhanced Segmentation and Dynamic Detection Recovery for Missed Detections in Dynamic Scenes

In the field of visual simultaneous localization and mapping (SLAM), real-time and precise environmental measurement is crucial for successful localization and mapping. In dynamic scenes, some SLAM systems that integrate semantic segmentation networks often overlook the need for rapid response capabilities on general devices. Correspondingly, the use of lightweight object detection networks might result in underutilization of valuable feature resources. This paper presents an enhanced SLAM system that integrates object detection in parallel with the frame construction process, achieving optimized speed and accuracy suitable for dynamic environments. First, the object detection model is improved to balance detection accuracy and inference speed, achieving 16.2% faster inference speeds for the object detection network at the cost of 0.5% reductions in detection accuracy. Then, depth information from RGB-D images is used to perform dynamic field segmentation on detected dynamic objects, enhancing feature preservation. Additionally, a recovery scheme for dynamic objects with missed detection is proposed to further improve accuracy. Finally, dense mapping of the scene is performed. FDR-SLAM achieves a reduction in absolute trajectory error (ATE) ranging from 89.3% to 96.1% on the TUM high-dynamic dataset, and from 89.9% to 98.8% on the Bonn high-dynamic dataset, compared to ORB-SLAM3. By comparing with state-of-the-art frameworks, FDR-SLAM achieves a balance between speed and accuracy.