Classification of military vehicles and weapons with drones using deep learning architectures with drones

In this study, a drone that can detect military vehicles and armed and unarmed people by making autonomous flights from the air was designed with the model trained using deep learning methods. To classify objects, 5 different transfer learning architectures such as ResNet50, InceptionV3, Xception, Mobilenet, EfficientNet were studied and performance comparisons were made. Additionally, the performances of YOLOv8 and YOLO-NAS architectures were compared for object detection. A new 6-class hybrid dataset was prepared for architectures, containing 4195 images of military and civilian vehicles, civilian and armed people. In order to prevent the models from overlearning, data multiplication methods such as random approximation, horizontal random shift, vertical random shift, horizontal rotation, rotation at a certain angle and horizontal shift were applied to the images in the data set. In the study, the most successful results in transfer learning architectures were obtained with the ResNet50 architecture with an accuracy of 0,9824 and an F1-score value of 0,9802. The highest scores in object detection were obtained with the YOLOv8 object detection algorithm as 0,961 mAP50 and 0,9116 F1-score. Since the ResNet50 architecture showed lower performance than YOLOv8 in the real-time operating tests performed on the designed UAV, the model created with YOLOv8 was added to the UAV. The data can be increased by adding more aerial images of military vehicles to the data set. Each class in the data set can be customized and divided into different classes.