Fig. 1. System framework diagram

Fig. 2. Improved YOLOv5s-seg network structure

Fig. 3. Two Ghost modules. (a) DFC attention; (b) GhostNetV2Bottleneck module; (c) GhostV2C3 module

Fig. 4. Schematic of the epipolar constraint. (a) Case of meeting the epipolar constraint; (b) case where the epipolar constraint is not met

Fig. 5. Removing dynamic feature points. (a) Semantic segmentation; (b) semantic segmentation removal; (c) pole constraint removal; (d) weighted geometric constraint removal

Fig. 6. Absolute trajectory error map, 3D trajectory error heat map, and relative pose error map. (a) fr3/walking_xyz sequence; (b) fr3/walking_rpt sequence; (c) fr3/sitting_static sequence

Fig. 7. Experiments using RGB-D cameras in a real dynamic scene. (a) Experiment scene; (b) experimental result

Fig. 8. Comparison of different algorithms in a real dynamic scene