Application of YOLOv7 and Transformer structures to small object (license plate) detection in complex traffic scenes

YOLO is a family of widely used one-stage object detectors, and YOLOv7 is the latest one of the series. In traffic application scenarios, such as night scenes, the images or videos often appear to be blurred, too dark or bright, or the distance might be too far away or too close. These situations may pose great challenges to the detection. In this paper, we examined the detection ability of YOLOv7 to small targets such as license plates in complex traffic scenarios by training and testing it on the specific dataset, and made changes to YOLOv7 with Transformer structures to see the performance of the modified YOLOv7 on the same task. We have confirmed that YOLOv7 performs very well in the task of license plate detection, and got accuracy significantly better than the object detectors of the previous generation. When we apply Transformer-style module to the network, we got lower FLOPs than the original YOLOv7, and successfully kept the accuracy at the same level.