Robust SLAM Algorithm Based on Semantic Information and Edge Consistency

To handle the performance degradation and the insufficient robustness of visual localization in dynamic environments, and to improve the created environment map, a robust simultaneous localization and mapping (SLAM) algorithm based on the semantic information and the edge consistency is proposed. Firstly, the semantic information of the environment is acquired by YOLOv3 algorithm, and the semantically dynamic-static segmentations of the image are obtained preliminarily. The consistency evaluation is conducted based on the distance transform errors and photometric errors of edges in the images to refine the dynamic-static area. Moreover, the dynamic regions are corrected by the connected component analysis and the loophole mending algorithm. The feature points in the non-dynamic regions are matched and the camera poses are optimized by minimizing the reprojection errors of feature points by the nonlinear optimization algorithm. The mapping keyframes are selected based on the covisibility of feature points and the areas of the dynamic-static regions. And the static environment map is created without the information of the dynamic objects. The experiment on the highly dynamic scenes in the public datasets show that the proposed method can distinguish the dynamic-static information accurately, and perform the precise localization and mapping in dynamic environments. Besides, the degradation of the positioning accuracy does not exist in the proposed method in the static environment.