Less is more: An effective method to extract object features for visual dynamic SLAM

Abstract

Visual Simultaneous Localization and Mapping (VSLAM) is an essential foundation in augmented reality (AR) and mobile robotics. Dynamic scenes in the real world are a main challenge for VSLAM because it contravenes the fundamental assumptions based on static environments. Joint pose optimization with dynamic object modeling and camera pose estimation is a novel approach. However, it is challenging to model the motion of both the camera and the dynamic object when they are moving simultaneously. In this paper, we propose an efficient feature extraction approach for modeling dynamic object motion. We describe the object comprehensively through a more optimal feature selection strategy, which improves the performance of object tracking and pose estimation. The proposed approach combines image gradients and feature point clustering on dynamic objects. In the back-end optimization stage, we introduce rigid constraints on the dynamic object to optimize the poses using the graph model and obtain a high accuracy. The experimental results on the KITTI datasets demonstrate that the performance of the proposed approach is efficient and accurate.