Searching for a needle in a haystack: Small Traffic Sign Detection in Large Images

A crucial component of driver assistance systems and self-driving automobiles is the identification of traffic signs. Traffic detection plays an important role in most sectors, like urban planning, transport management, and driving systems. The main issue is that in photographs taken from the road, traffic signs appear substantially smaller. Only 1% to 2% of the entire image area is covered by it. Therefore, it can be difficult to see a tiny traffic sign in an image with a wide background of items with similar shapes. As a result, to detect small traffic, it is suggested to use YOLOv3 network layers. This helps to raise the mean average precision and recall percentages. In order to get the best anchor set for the dataset, this work also provides an anchor box algorithm that will make use of bounding box dimension density. This lowers the average miss rate and false positive rate. This research paper provides a suitable approach to detecting minute traffic signals in a large image. The detection process goes through various steps, including managing the large image to an appropriate size to reduce the complexity and enhance the picture quality. Various algorithms are applied to it, like YOLO and CNN. This proposed approach is demonstrated by various experimental results. The suggested method is tested against Swedish and German traffic sign datasets.