YOLOP‐FCD: A Multi‐Task Visual Perception Network with Progressive Feature Decoupling and Attention Synergy for Autonomous Driving

ABSTRACT

Panoptic driving perception, which unifies object detection, lane line recognition and drivable area segmentation, is a cornerstone of autonomous driving. Multi‐task learning (MTL) is a natural paradigm to tackle these heterogeneous tasks jointly; however, conventional MTL frameworks often suffer from negative transfer, where conflicting task objectives lead to suboptimal feature representations and degraded accuracy. Such deficiencies pose safety risks in real‐world autonomous systems. In this paper, we propose YOLOP‐FCD, a novel multi‐task visual perception network that mitigates negative transfer through two key innovations. First, a progressive layered extraction module explicitly decouples task‐specific features while enabling selective feature sharing in a hierarchical manner, balancing common knowledge transfer with task isolation. Second, a synergistic multi‐attention transformer block integrates pixel‐wise, channel‐wise and spatial attention to enhance feature representation and cross‐task collaboration. Extensive experiments on the BDD100K dataset demonstrate that YOLOP‐FCD achieves competitive performance and a more balanced trade‐off among perception tasks. Ablation studies further validate the effectiveness of the proposed modules in alleviating feature interference and improving task synergy for autonomous driving perception.