Optimizing Lane Detection in Autonomous Vehicles Using Cascading Attention Mechanisms in DeepLabv3+

The high rate of traffic accidents caused by human error highlights the urgent need for reliable lane detection in autonomous vehicles. Traditional segmentation models, such as U-Net, struggle to capture multi-scale contextual features, leading to inaccurate recognition of narrow or visually ambiguous lane markings. This study introduces an enhanced DeepLabv3+ architecture, augmented with cascading attention mechanisms, Convolutional Block Attention Module (CBAM), Efficient Channel Attention (ECA), and Squeeze-and-Excitation (SE), to improve feature extraction and boundary precision. The proposed method addresses limitations in existing models by leveraging these attention modules to better capture contextual information at different scales. A dataset consisting of 374 annotated road images from Makassar, Indonesia, was used for training and evaluation. The model achieved a mean Intersection over Union (IoU) of 97.34% and a Root Mean Square Error (RMSE) of 0.0377 in tire-to-lane distance estimation, outperforming traditional architectures. These results demonstrate that the proposed framework provides robust, real-time lane detection, making it highly suitable for autonomous vehicle navigation in dynamic and complex urban environments.