EdgeUNet: Edge-Guided Multi-Loss Network for Drivable Area and Lane Segmentation in Autonomous Vehicles
Xing Sheng, Junzhe Zhang, Zhen Wang, Zongtao Duan
Abstract
The perception system is a critical element of autonomous driving, where real-time and accurate segmentation of drivable areas and lanes is essential for intelligent decision-making during vehicle operation. Current approaches primarily focus on minimizing background interference in images, often overlooking the importance of edge information. In response, this paper introduces the edge-guided multi-loss network (EdgeUNet) designed for drivable area and lane segmentation. EdgeUNet employs an encoder for feature extraction and a decoder specifically tailored for the segmentation tasks. The decoder incorporates a novel feature fusion module (FFM), multi-scale feature aggregation module (MFSA), edge extraction module (EEM), and edge-aware optimization module (EAO), facilitating efficient extraction and supervision through edge information. Our model demonstrates superior performance on the Berkeley deep drive (BDD100K) dataset, achieving state-of-the-art results with 99.3% mean pixel accuracy (mPA) and 54.8% mean intersection-over-union (MIoU) in the lane detection task. Additionally, ablation studies conducted on the TuSimple and KITTI datasets further validate the effectiveness and generalizability of EdgeUNet.