ArticuBEVSeg: Road Semantic Understanding and its Application in Bird's Eye View From Panoramic Vision System of Long Combination Vehicles

Abstract

Long combination vehicle (LCV) plays a vital role in modern urban transportation, enhancing traffic efficiency and alleviating congestion. However, the inherent unique characteristics of LCV, including articulation joints and extended lengths, pose challenges to its operational reliability, compromise road safety, and moreover significantly hinder advancements in autonomous driving. To address some of these challenges, we propose ArticuBEVSeg, an innovative framework for road semantic understanding in bird's eye view (BEV) using panoramic vision systems of LCVs. ArticuBEVSeg incorporates an implicit temporal alignment mechanism without ego-motion data while seamlessly and directly integrating spatial information from distorted fisheye images with time-varying extrinsics caused by articulations among carbodies. Furthermore, based on ArticuBEVSeg's segmentation results, a lane positioning algorithm for LCVs that bypasses post-processing instance clustering was proposed, enabling robust lateral control feedback for LCVs in dynamic road scenes. To validate the proposed algorithms, we collected the first dataset featuring panoramic fisheye images from real-driving scenarios of articulated LCVs. Experimental results verifies the effectiveness of the proposed algorithms in realizing road semantic understanding while being robust against road scene variation and articulation angle accuracy. The proposed algorithms in this study introduce a flexible BEV perception paradigm adaptable to diverse BEV perspective encoders and transportation means, promoting automated driving of special vehicles like LCVs in urban traffic.