Research on dynamic scene SLAM algorithm incorporating semantic information

• 1、School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, 100876

Abstract：Simultaneous localization and mapping (SLAM) technology is a key technology applied to robot navigation, unmanned driving and other fields. The traditional SLAM framework usually takes the static scene as an a priori assumption, but the dynamic objects that may exist in the actual application scene will cause serious interference to the system operation, resulting in the drift of the position estimation and the distortion of the map construction. For the impact of dynamic objects on the accuracy of SLAM system in real scenes, we propose a dynamic feature point rejection algorithm fusing semantic information. The algorithm models the semantic a priori information output from semantic segmentation network and geometric constraints as a joint dynamic probability as a judgment threshold to reject the dynamic feature points in the image and avoid the interference of dynamic objects on the system. Experimental results show that the algorithm significantly reduces the influence of dynamic objects on the position estimation in highly dynamic scenes and significantly improves the localization accuracy of the SLAM system.

Keywords： Computer application techniques simultaneous localization and map building semantic information deep learning dynamic scenes