Development of a BEV Lane Detection Model Using a Dual Attention-based Decoder

Lane detection is critical for autonomous driving systems in understanding the surrounding road environment. While most prior studies have focused on pixel-wise lane segmentation in the front-view image space, their performance is often limited by structural constraints such as perspective distortion and a narrow field of view. To address these limitations, we propose a novel lane detection method that operates in the bird’s-eye view (BEV) image space, providing enhanced spatial understanding. The main contributions of this study are as follows: 1) We design a simple yet effective deep neural network architecture that predicts BEV lane segmentation from a single front-view camera image. 2) We incorporate a dual-attention module to improve the representation of BEV feature maps across both channel and spatial dimensions. 3) We converted the OpenLane dataset, containing 3D lane annotations, into a BEV lane segmentation dataset, on which we evaluate the effectiveness of our approach. Experimental results demonstrate that the proposed model with dual-attention improves mIoU score by 9.3% compared to a baseline architecture.