Enhancement of UFLD by improving global dependencies

<p dir="ltr">Lane detection is a vital component of autonomous driving technology. It identifies the positions and boundaries of lane lines in images captured by onboard cameras, facilitating key autonomous driving functions and enhancing road safety. Deep learning-based methods dominate lane detection task, with the Ultra Fast Lane Detection (UFLD) model being one of the most well-known recent approaches. Unlike other pixel segmentation-based methods, UFLD distinguishes itself through its row-anchor detection approach, enabling rapid processing speed. However, UFLD encounters challenges related to detection accuracy in various complex scenarios. An analysis of UFLD architecture revealed that insufficient global dependencies limit its performance and generalizability in complex scenarios. To address this limitation, this paper proposed a non-local UFLD model, which strengthens the global dependencies by integrating non-local blocks. Additionally, an auxiliary Lane Intersection over Union (LIoU) loss function is introduced to refine the model’s ability to accurately detect the position and shape of the lane lines. Experimental results on the CULane dataset show that non-local UFLD surpasses original UFLD in detection accuracy across most scenarios while maintaining high detection speed.</p>