A novel automatic target reporting method with STC-YOLO for videos

Live-fire drill is an essential component of military training. However, target reporting systems face many problems, including high costs, low real-time judgement, and susceptibility to the environment. In recent years, deep learning have been applied to automatic target reporting systems to address these issues. These methods face the challenge of detecting small objects. Besides, current deep learning-based methods are primarily applied to static image detection. For dynamic detection of video, they can be disturbed by the movement of the target paper or camera when recognizing new bullet holes. Therefore, this paper proposes a novel target reporting method for videos. Firstly, for small object detection, this paper proposes STC-YOLO, where we simplify the backbone network, design WXFPN as the neck network, and introduce the SPD-Conv module and the SIoU loss function. Secondly, for videos, we combine the BoT-SORT tracking algorithm to achieve dynamic recognition of bullet holes. Thirdly, we design a bullet hole scoring algorithm to achieve the complete target reporting functionality. This method achieves high accuracy in bullet hole detection, with STCYOLO achieving scores of 0.996, 0.98, and 0.99 in precision, recall, and mAP50, respectively, on the test set. At the same time, this method achieves accurate bullet hole scoring, with a scoring accuracy of 92.8%. Experimental results proves the validity of this method.