Visual SLAM Based on Semantic Segmentation and Geometric Constraints for Dynamic Indoor Environments

Abstract

Simultaneous Localization and Mapping (SLAM), as the core technology of mobile robots, has received more and more attention in recent years. However, most of existing visual SLAM algorithms do not consider impact of dynamic objects on localization accuracy of visual SLAM. In view of this, based on ORB_SLAM2 algorithm, this paper proposes a dynamic feature point filtering algorithm for semantic segmentation and motion consistency detection with grid homogenization. Aiming at the problem that feature points extracted by ORB_SLAM2 are easy to gather in areas with rich texture information, and feature point extraction rate is low in weak texture information areas, to solve these problems a grid-based feature point homogenization algorithm is proposed. Through statistical analysis the grayscale value of current node to calculate a threshold for feature point extraction, and then the extracted feature points are retained by Harris response value and the Euclidean distance. In addition, aiming at the impact of dynamic targets on the SLAM system, a semantic segmentation network and a motion consistency detection algorithm are combined to improve the robustness of the system. The prior dynamic in-formation is obtained through semantic segmentation, and dynamic feature points are filtered out in combination with the motion consistency detection algorithm proposed in this paper. Then, a semantic map is constructed by combining semantic information. Finally, the improved algorithm was tested on Oxford data set and TUM data set. The experimental results show that improved algorithm can effectively improve positioning accuracy in dynamic environments, and constructed semantic map is rich in semantic information.