A Visual-Inertial Dynamic Object Tracking SLAM Tightly Coupled System

Abstract

Most existing dynamic simultaneous localization and mapping (SLAM) systems simply remove dynamic objects, resulting in the loss of dynamic object motion information that helps the system itself localize and navigate. Aiming at the above problems, a visual inertial (VI) dynamic object tracking SLAM tightly coupled system is proposed. First, a method for classifying dynamic and static feature points is proposed to solve the problem that dynamic objects cannot be extracted by learning-based or vision-only methods. Second, the system accurately extracts and tracks dynamic rigid-body objects in the scene through a combination of visual information, inertial measurement unit (IMU) data, and associated matching point labels, without modeling the shape or geometry of the dynamic objects. Subsequently, a novel pipeline is proposed to jointly optimize the camera poses, trajectories of dynamic objects, the maps of the environment, and IMU information, improving the self-localization performance of the system. Finally, the performance of the system is tested on indoor and outdoor datasets and real dynamic scenes. Experimental results show that the proposed system achieves improved performance in self-localization and dynamic object tracking compared to other state-of-the-art methods. Especially under adverse conditions of partial occlusion and far-field of view, the system can still track dynamic objects robustly and accurately to optimize self-localization performance.