Enhanced Initialization for Track-before-Detection based Multibody Motion Segmentation from a Moving Camera

Abstract

Vision-based motion segmentation provides key information for dynamic scene understanding and decision making in autonomous navigation. In this paper, we propose an enhanced initialization strategy which is tightly coupled to a multibody Structure from Motion (SfM) segmentation employing a monocular camera. The method relies on epipolar geometry, RANSAC formulation and motion estimation for segmenting ego-motion and eoru-motions. The proposed strategy is intended to enhance the initialization procedure employed in [1] obtaining a 50 times speed-up factor. This result is used as input for the Track before Detection (TbD) methodology which efficiently simplifies the existing multibody Structure from Motion implementation approaches. An evaluation using images from a publicly available dataset with dynamic traffic scenarios and large camera motions confirms the effectiveness of the method.