Feature removal of movable objects in visual SLAM

Abstract

Simultaneous localization and mapping (SLAM) in dynamic environments is a challenging problem in robotics navigation. While leading SLAM algorithms have become very robust in static environments, moving objects in dynamic environments can cause multiple landmarks to be created in the map, decreasing the accuracy of localization in dynamic environments. To address this problem, we propose a system composed of an object detection and classification system, YOLOv3, and a visual SLAM system, ORB-SLAM2. In our system, objects defined as movable are detected from the current frame and ORB features that correspond to the detected objects are removed. With our system, we conduct experiments in multiple environments and compare the resulting trajectories with those of the original ORB-SLAM2. Our results show that in dynamic environments our system can successfully remove features that belong to movable objects and produce a more accurate trajectory.