Técnicas de pré-processamento utilizando YOLOP para detecção de faixas em ambientes de direção autônoma

Abstract

The field of autonomous driving has garnered increasing interest in solutions that combine high reliability with real-time inference capabilities. The diversity of real-world scenarios, including variations in lighting and weather conditions, imposes additional challenges on tasks such as object detection and semantic segmentation. This study investigates the integration of lighting correction techniques into the YOLOP system, a single-stage deep learning algorithm designed for lane detection on highways. The AWB, SCL-LLE, and Wavenet techniques were evaluated using images from the VIL-100 and CULane datasets. Experimental results revealed that the SCL-LLE model achieved the best performance, highlighting that leveraging semantic information during lighting correction positively impacts lane detection, aligning with YOLOP’s approach to utilizing semantic features.