Feature-Based SLAM and Moving Object Detection and Tracking with Ego-Motion Compensation

Abstract

SLAM in real dynamic environments is the key component to making the robotic system more useful and intelligent. From now on, most researchers have considered the world as static status. However, avoiding dynamic objects in the real world is inevitable, which deteriorates both pose estimation and mapping process. In this paper, we connect the SLAM with DATMO module, which includes detecting moving object and tracking them. For detecting moving object, the current pose of the egorobot is required to be compensated to exclude its own movement. However, it means that the current pose is needed before the SLAM's pose estimation process. To solve this problem, we exploit the Kalman Filter based predictor to offer pose value to DATMO module. In addition, by including some parameters memorizing the temporal moving status in tracking, we can cope with the objects which are temporarily in static status. All of our systems is built on ROS framework and we evaluated our method on KITTI dataset, which is mainly composed of the town or road scenes. The results show that our method is especially robust to heavy traffic conditions and also stable in other relatively static environments.