A Higher-Accuracy Color- and Gray-Image Training Method for Road-Object Detection with YOLO

Abstract

This paper proposes a higher-accuracy color- and gray-image training method for road-object detection with You Only Look Once (YOLO). Image pre-processing is performed before feeding images to a network for object detection, i.e., color-to-gray conversion and edge detection. Then, the original images, the gray version of the original images, and the edge images are fed to the main network. YOLO version 5 (YOLOv5) was used as a backbone network. The model was trained by using custom traffic dataset which consists of 738 training images and 185 validating images, and they are separated into 7 classes. The proposed method achieved a mAP of 44.97% when performing a validation by using both RGB and grayscale images. It is higher than that of the conventional training method using only original images. The results also confirm that higher accuracy can be achieved even for night vision images. The proposed method could serve for applications which be used in various factors and environments.