Detection of military threat elements using deep learning techniques

Throughout history, one of the key drivers behind technological advancements has been the security concerns of nations in the face of military threats. The nature of these threats and the security measures taken to counter them vary depending on time, location, and the prevailing conditions of the period. Nevertheless, regardless of the circumstances, the first step in defending against military threats is their detection. For this reason, the identification of military threats constitutes a critical component of national defense strategies. Today, with the aid of deep learning techniques, it has become possible to detect potential threats through camera footage. Although there are studies in the literature addressing this issue, two major limitations commonly stand out: insufficient processing speed and lack of portability. Considering these two shortcomings, this thesis aims to develop a deep learning-based model capable of detecting military threats both rapidly and in a portable manner. Within the scope of this study, a military threat detection system based on deep learning techniques was designed using the YOLO (You Only Look Once) algorithm on the NVIDIA Jetson Nano development board. For this purpose, a dataset containing 1,795 images and consisting of the classes tank, military truck, and APC-IFV (Armored Personnel Carrier - Infantry Fighting Vehicle) was obtained from the Kaggle platform. Data augmentation techniques were applied to improve the overall performance of the model. An object detection model based on the YOLOv5n architecture was developed for training, and optimization was performed using TensorRT to enable real-time performance. The optimized model achieved a mean Average Precision (mAP) score of 89% across all classes. In order to evaluate the real-time performance of the model, it was tested on a video containing a military convoy, yielding a result of 22.1 frames per second (FPS). Furthermore, through the use of a multi-threaded video processing approach, this performance was increased to an average of 26.42 FPS.