AttBi-ResU: A Smart 3D Vision-Based Lane Detection System for Autonomous Vehicles

To enhance driving safety and minimize the risk of collisions, autonomous vehicles rely on lane-detection systems capable of issuing timely warnings during unexpected lane departures. Current methods often suffer from complex scenes, fluctuating illumination, high false-positive rates, and imprecise localization. To overcome these constraints, we present a novel three-stage vision-based pipeline. In the first stage, raw RGB images sourced from public repositories are converted into grayscale. This process leverages a bilateral entropy–based adaptive histogram equalization module, designed to enhance image contrast while effectively mitigating noise. An attention-augmented Bidirectional Long Short-Term Memory (Bi-LSTM) feature extractor feeds a residual-dilation U-Net, enabling precise spatiotemporal segmentation of lane regions. Finally, to minimize training loss, the model’s hyperparameters are adjusted using an enhanced Krill Herd Optimization (KHO) process.