Ai-based target detection system for ground combat vehicles

Modern military intelligence systems are continuously evolving. Particularly in environments where ground forces operate on the battlefield, the classification of combat vehicles has become critically important. Identifying the capabilities and vulnerabilities of various types of combat vehicles is essential for developing effective tactics and strategies. The use of artificial intelligence in the classification of military vehicles, through image processing, pattern recognition, and deep learning technologies, has become more feasible compared to the past. "You Only Look Once" (YOLO) algorithm provides significant advantages in object detection and image classification. YOLO possesses the ability to rapidly detect targets, enabling real-time object predictions. The algorithm's contribution to detection and evaluation processes with high accuracy offers substantial support to military personnel on the battlefield. Its ability to produce accurate results with low background errors and to understand general object representation further enhances the effectiveness of this method. In this thesis, the YOLO algorithm has been employed for the detection of combat vehicles. The YOLOv8m model has been selected due to its suitability for mobile applications and its demonstrated success in performance evaluations. The performance has been further enhanced with the addition of the squeeze and excitation block to the model. The study focuses on detecting detailed characteristics of targets using images obtained through electro-optical systems and emphasizes the decision support process. The objective of this thesis is to contribute to the identification and definition of the fundamental components of target management, thereby strengthening decision support processes. The process encompasses the development of a deep learning-based "Target Detection and Identification" (TDI) system, which includes specialized data collection, pre-processing, segmentation, basic feature extraction, classification, advanced feature extraction, and the provision of recommendations to users through decision support matrices. The developed model offers substantial advantages in image processing and object detection. Moreover, with the designed model, the passive identification of combat vehicles can be achieved using existing electro-optical devices, without requiring additional hardware, by estimating distances while keeping the observer's position concealed. The model provides real-time predictions with high accuracy (mAP 83.6%), thereby assisting in informed decision-making and minimizing risks.