This study investigates the efficiency of the You Only Look Once (YOLO) object detection algorithm for real-time traffic light detection under challenging real-world situations. Specifically, the study aims to quantify the performance and degradation in mean average precision (mAP) and measure the average inference speed (in frames per second) across different adverse weather classifications: rain, snow, and fog, while comparing against a dataset of clear skies weather. By combining large diverse datasets found in publicly available domains, a YOLO variant will be trained and evaluated against another separate dataset of images and videos of traffic lights. The findings will contribute to providing necessary benchmarks for evaluating the reliability in upcoming technological systems such as advanced driving assistance and autonomous vehicle systems.
