Pallid Faded Crosswalk Detection Using YOLO Algorithm

Abstract

Crosswalks deteriorate and fade from everyday use worldwide, which can lead to accidents between vehicles and pedestrians crossing the crosswalks. The machine learning techniques have the potential to extract patterns from pallid and faded crosswalks by creating models from images of crosswalks at all levels of clarity and images without crosswalks. we have chosen to use YOLOv8 as the model for the crosswalk detection system at all levels of clarity. This model is selected because it demonstrates good efficiency in object detection in images and can process data quickly. We have compiled a dataset of crosswalks at various levels of clarity by driving and collecting data in the Bangkok metropolitan area and its outskirts, which contain zebra crossings of different clarity levels. Furthermore, we have developed a system to classify crosswalk images into three levels of clarity: clear, damaged and faded to a low extent, and heavily damaged and faded. This system also includes images without crosswalks. These images are used to create datasets for zebra crossings at all levels of clarity, following the proportions of training, validation, and testing data, which are used to build a model for detecting crosswalks at all levels of clarity. And after testing the model with the Test dataset, the model achieved an mAP50-95 score of 77.3% and an mAP50 score of 96.2%. As for Precision, it reached 94%, and Recall reached 93%. This crosswalk detection system at all levels of clarity will serve as an assistive component in developing the autonomous car's cruise control or braking system when approaching zebra crossings of any clarity level. It will contribute to promoting safe driving of autonomous vehicles in the vicinity of crosswalks in all clarity levels.