Correção de Iluminação usando YOLOP para Detecção de Faixas em Ambientes de Direção Autônoma

This work investigates the impact of illumination correction on lane detection in autonomous driving scenarios. We employ YOLOP, a deep learning-based architecture capable of real-time drivable area segmentation and lane detection. Three preprocessing techniques were evaluated: AWB, SCL-LLE, and Wavenet. Experiments were conducted using the VIL-100 and CULane datasets, with standard evaluation metrics including mP, mR, mIoU, and mAP. While AWB showed limitations under adverse lighting conditions and Wavenet delivered moderate results, SCL-LLE stood out for its semantic-aware correction, yielding the best overall performance, particularly on the VIL-100 dataset. We conclude that techniques leveraging semantic information significantly enhance YOLOP's robustness and detection accuracy under varying lighting. The proposed approach proves to be a viable strategy to improve visual perception in embedded systems for autonomous vehicles