Efficient MobileNet Backbone and Cascade Fusion in CFRNet for Robust Lane Detection

Robust lane detection remains a critical challenge for autonomous driving systems, particularly under complex road conditions and real-time computational constraints. We adapt the Cascade Fusion Network (CFRNet), originally designed for road extraction in high-resolution remote sensing imagery, to the task of road lane segmentation. Our approach employs a lightweight MobileNet-based backbone for real-time feature extraction and a cascade of sub-backbones, each equipped with a triple-level adaptive feature fusion (TAFF) module, to integrate multiscale spatial and contextual cues. Each cascade stage progressively refines multiscale spatial details and contextual dependencies, enabling robust identification of lane markings under occlusion, low illumination, and sparse road topology. Extensive experiments on the TuSimple benchmark demonstrate that CFRNet-Lane achieves state-of-the-art accuracy (98.28 % accuracy) with a 2.52 % reduction in false negatives compared to ResNet-based counterparts, while operating at 60 FPS on embedded hardware. The proposed model strikes an optimal balance between precision, adaptability to challenging environments, and computational efficiency, making it viable for deployment in resource-constrained automotive systems.