Towards Real-time Semantic SLAM Based on Undesired Regions Segmentation

Visual SLAM (vSLAM) is an essential technology for providing robots with position and environment information in unknown environments. But the problem of robustness has not been well solved. For example, the features near the skyline will aggravate the scale drift of monocular SLAM systems, and the existence of challenging dynamic objects violates the static world assumption. In this paper, a robust vSLAM method to remove three undesired regions of the sky, skyline and potentially dynamic objects, building on ORB-SLAM2, is proposed. We incorporate a real-time semantic segmentation network into a vSLAM system to quickly obtain accurate semantic information to eliminate the influence of undesired regions on localization and mapping. In addition, for real-time efficiency, we put the semantic segmentation network in a separate thread and perform semantic segmentation only on keyframes. We conduct monocular and stereo SLAM experiments in 11 sequences (00–10) of the KITTI dataset. The evaluations show that our method achieves a 44.44% accuracy improvement over the baseline ORB-SLAM2 in the best case, and alleviates the scale drift in the monocular case. Compared with DynaSLAM, the advanced dynamic semantic SLAM, our method is slightly more accurate and much faster.