Advancing Road Lane Detection in Autonomous Driving through Multistage Attention Network

Developing reliable autonomous systems requires road lane segmentation models that can mimic human perception without the associated errors. This paper introduces MARes UNet, a model previously used for segmenting highly detailed satellite images, which is now applied to road lane segmentation. Traditional U-Net structures rely heavily on long skip connections between the encoder and decoder and perform poorly under challenging conditions such as extreme weather, traffic congestion, and low light. Although the dot-product attention mechanism performs better, it also leads to scaling issues. MARes-UNet incorporates a Linear Attention Module (LAM) within the skip connections to improve global context understanding and eliminate the scaling problem. Evaluations of the effectiveness of the model were conducted using the TuSimple dataset, which includes difficult weather conditions. When compared to many cutting-edge models, MARes-UNet showed greater accuracy and a lower False Positive Rate, indicating its potential for reliable road lane segmentation in real-world applications.