SSF-MOS: Semantic Scene Flow Assisted Moving Object Segmentation for Autonomous Vehicles
Tao Song, Yunhao Liu, Ziying Yao, Xinkai Wu
Abstract
Detecting moving objects in dynamic environments precisely is essential in autonomous driving. Existing object detection methods using point clouds have difficulties to distinguish moving and static objects in dynamic environments. Motivated by the optical flow method widely used in image-based dynamic object perception, we propose SSF-MOS (Semantic Scene Flow assisted Moving Object Segmentation), a unified framework that incorporates semantic information and ego-motion estimation in moving object segmentation. SSF-MOS first detects and excludes absolutely static objects, such as poles and roads, by applying the semantic segmentation method. Subsequently, the proposed semantic scene flow estimation method computes the motion vectors between consecutive point clouds and predicts the motion state (moving or static) of each point. Furthermore, SSF-MOS calibrates results of moving points by considering the ego-motion of autonomous vehicles. We directly introduce semantic information in the decoupled framework for more accurate results and convenience of upgrades. The extensive experiments show that the proposed SSF-MOS achieves competitive performance of 0.701 mIOU compared with other state-of-the-art methods on the public dataset SemanticKITTI.