Improvement of Small Object Detection for Road Vehicles Based on DINO

The task of vehicle object detection on expressway has always been a topic of concern. At present, YOLO is mainly used to achieve high image reasoning speed, but the accuracy is not ideal. Bounding box regression (BBR) has been widely used in object detection and instance segmentation,which is an important step in object localization. Therefore, the loss function used to return the object position is crucial to the regression speed of the bounding box and the impact on the positioning accuracy of the target position. Therefore, a new loss function is proposed. The process is to model the bounding box as a 2D Gaussian distribution, and then propose a new metric called normalized Wasserstein distance (NWD), and calculate the similarity between them through the corresponding Gaussian distribution. This method can be used to detect objects in aerial images with the size of only a few pixels. We apply this method to DINO, and the detection accuracy of small vehicles in the distance is improved by approximately 2.1AP.