Visible light small object detection based on YOLOv5

Abstract

The traditional object detection algorithm is difficult to extract its characteristic information due to its own features such as low resolution and small coverage area of small objects, resulting in the inability to achieve effective and reliable recognition accuracy. Aiming at the problem of small object detection, this paper proposes a method of visible light small object detection based on deep learning YOLOv5 algorithm. First of all, a total of 4000 visible light small object dataset is created at noon and low light under the background of sunny and cloudy weather, and then YOLOv5 is used for training, of which the mAP@0.5 of 100 and 200 times are trained to reach about 95% and 96%, respectively. Finally, the 500 pure sky background visible light small object images outside the dataset are tested using the trained model, and the recognition rate in sunny weather reached 99%. However, in cloudy weather, due to the interference of clouds, false detection and missed detection occur, and the recognition rate is about 97%. For the phenomenon of false detection, the moving object detection algorithm are combined to exclude. First of all, a small amount of large particles of pretzel noise is added, combined with the moving object detection algorithm, the motion trajectory is plotted for the continuously moving visible small objects, so as to exclude the noise that is far away from the motion trajectory, the coarse filtration rate reaches 79.5%, and the remaining target point collection is further filtered out by DBSCAN clustering algorithm, and the noise filtering rate can reach 100%.