Research on SLAM System incorporating Weakly-Supervised Learning in Dynamic Environment

Abstract

Based on Simultaneous Localization and Mapping (SLAM) and deep learning, a SLAM system incorporating weak supervised learning semantic segmentation is proposed for navigation and positioning of mobile robots in dynamic environment. After obtaining the image from the camera, the weakly supervised learning semantic segmentation network SEC was used to segment the semantic information, and then the obtained semantic information was used to eliminate the dynamic feature points, so as to improve the accuracy of pose estimation of the system. Compared to traditional orb-slam2, this article shows an overall improvement in performance on publicly available high dynamic sequence data sets. The results show that after improvement, both absolute error, relative drift and rotational drift are significantly reduced, which proves that compared with orb-slam2, this method can improve the accuracy of pose estimation in dynamic environment.