Real-Time Military Tank Detection Using YOLOv5 Implemented on Raspberry Pi

Military target detection is an essential step to improve battlefield situation generation, surveillance, and command decision-making. In this paper, we utilized YOLOv5 as a popular deep learning model for tank detection in automated war operations. Due to a limited availability of domain-specific datasets, we collected a real dataset and then customized it with some open source images from different military tanks, light and weather conditions. Tank instances were manually annotated under the guidance of military experts. We evaluated the out-of-the-box performance of three detectors of YOLO (YOLOv5s, YOLOv5m and YOLOv5l) on our dataset and compared them in terms of accuracy and run-time. In addition, transfer learning is applied to overcome the inadequate data problem in military object detection. We also investigated the effect of data augmentation on the detection accuracy. The model YOLO5vl can got the maximum average precision (98.5%) compared with two other detectors. Run time of the detectors was compared with an implementation of Raspberry Pi which was connected to camera for data collection. YOLO5vs can recognize the military tanks in a captured RGB image with size 614×614 pixels at 7.9 ms per image. This results show that the applicability of YOLO5vs for the real-time military tank detection.