End2end vehicle multitask perception in adverse weather

In the research of autonomous driving technology, due to the lack of datasets for various extreme weather conditions, autonomous driving perception in adverse weather is a challenge. To address this problem, this paper introduces an end-to-end multi-task perception system that combines labeled supervised learning and unsupervised domain adaptive learning for bad weather. The key innovations of this system include: a multitask learning framework that simultaneously handles object detection, lane line detection, and drivable area detection, improving both efficiency and cost-effectiveness for autonomous driving in complex environments; a domain adaptation strategy using unlabeled data for adverse weather, which enables the system to perform robustly without requiring specific labels for harsh weather conditions; the system has strong generalization ability, demonstrated by achieving an prediction mAP of 83.86%, a drivable area mIoU of 91.59%, and lane detection accuracy of 83.9% on the BDD100K dataset, as well as an mAP of 74.85% on the Cityscapes fog dataset without additional training, highlighting its effectiveness in unseen, adverse conditions. The scalable and generalized solution provided in this paper can achieve high-performance navigation in various extreme environments. By combining supervised and unsupervised learning techniques, this model can not only cope with severe weather but also further generalize to unseen scenarios.