Abstract
Lane departure detection is a crucial task in advanced driver assistance systems (ADAS) and autonomous driving, aimed at reducing accidents caused by unintentional road deviation. This study proposes a modified U-Net architecture enhanced with Feature Cross Attention (FCA) to improve lane departure anomaly detection. The objective is to enhance spatial sensitivity and context awareness in segmentation, especially under challenging driving conditions such as occlusions, poor lighting, and distorted lane geometry. The materials used include the publicly available Comma 2k19 LD dataset, comprising 2,000 manually annotated frames extracted from highway driving scenarios. Each frame includes synchronized video and driving telemetry, offering diverse visual conditions. Preprocessing steps include resizing, normalization, and annotation conversion to binary masks. An anomaly is defined as a spatial deviation threshold between predicted and ground-truth lane boundaries. The proposed method incorporates FCA at the bottleneck and decoder levels of the U-Net architecture. Evaluation was performed using Intersection over Union (IoU), Pixel Accuracy, and threshold-based anomaly criteria. The model achieved 99.19% Pixel Accuracy and 98.47% IoU, outperforming the baseline U-Net (97.56% and 97.46%, respectively). Visual results showed improved detection of subtle lane shifts. A confusion matrix generated over 210 validation images demonstrated perfect classification of normal and anomalous cases. These results confirm that FCA integration enhances segmentation precision and anomaly sensitivity. The approach is suitable for real-time deployment in autonomous systems. Future research may focus on temporal integration, lightweight optimization for embedded devices, and the extension of the framework to multi-lane or urban traffic environments.