Robust Indoor Visual-Inertial SLAM with Pedestrian Detection

Abstract

Most Simultaneous Localization and Mapping (SLAM) systems are designed based on the scene rigidity assumption, which limits the applications of SLAM systems in highly dynamic environments. In addition, a single sensor cannot meet the needs of complex environments, and the fusion of visual and inertial sensors that can provide robust and accurate 6DOF pose estimation has become popular in robotics. In this work, a novel visual-inertial SLAM system towards dynamic indoor scenarios is proposed. Our approach is built on ORB-SLAM3, and a semantic segmentation thread is added to detect pedestrians. Feature points on people are avoided extracting to improve the localization accuracy in dynamic environments. The parallel tracking thread and segmentation thread can improve the real-time performance of our algorithm. We test our system both on TUM-VI dataset and in real-world environment, and the results show that our system can achieve better robustness and accuracy than ORB-SLAM3 in populated environments.