Construction of Semantic Point Cloud Based on Dynamic Object Recognition

Simultaneous localization and mapping (SLAM) is considered to be a fundamental function of intelligent mobile robots. However, some V-SLAM algorithms lack robustness in dynamic environments for the reason of the strong assumption of static world. Meanwhile, these algorithms are relatively simple to map dynamic objects, which generates error images and leads to map clutter. In this paper, we develop a semantic RGBD SLAM system. The system uses object detection and priori semantics, for high-precision localization and accurate scene reconstruction. Firstly, we introduce a dynamic object detection method that combines geometric computation and semantics. Secondly, by using detection and semantics, dynamic feature points are removed to optimize camera poses for improving the accuracy of trajectory estimation and reconstructing dynamic dense point clouds. Finally, we evaluate the SLAM system in public datasets environments. The results show that our system can achieve precise localization in dynamic environments while building a clear dynamic semantic point cloud map.