Implementation of aircraft detection and tracking system in real time on embedded systems

This thesis aims to develop an embedded system that will enable Unmanned Aerial Vehicles (UAVs) to gain pilot-level decision-making and target tracking capabilities in various combat scenarios. In this context, the integration of embedded systems and artificial intelligence models was addressed and it was aimed to provide target detection and tracking operations in real time. In the study, it was aimed to obtain a system with low latency by using technologies such as Python, OpenCV and YOLO. The system obtains a live image with a high-resolution camera placed on the nose of the UAV, processes this image on the embedded system, detects the target aircraft and starts the tracking process. The YOLO model was preferred in this project because it provides high accuracy and speed. After the location and orientation of the target aircraft are detected, the tracking process is continued with control signals sent to the flight computer. Thanks to this technology, dangerous scenarios that fighter pilots are exposed to can be managed automatically without the need for human intervention, which offers significant advantages in terms of safety and operational costs. Within the scope of the thesis's experimental studies, results obtained through various simulation software and virtual testing environments have demonstrated that the system performs successfully in terms of both accuracy and stability. In particular, the YOLOv8n model achieved an accuracy of 90.79% mAP@0.5 during training, indicating high performance in detecting fixed-wing unmanned aerial vehicles (UAVs). This level of accuracy places YOLOv8n among the most effective models optimized for embedded systems. In real-time tests conducted on a Jetson Nano platform, an average of 17.5 FPS was achieved, confirming the system's suitability in terms of both speed and detection accuracy. From the perspective of tracking algorithms, while the classical Kalman Filter provided adequate performance in simpler scenarios due to its low computational cost, it proved insufficient in complex maneuvers and nonlinear flight dynamics. In contrast, the use of the Extended Kalman Filter (EKF) resulted in high tracking stability in tests conducted on both desktop and Jetson Nano platforms, significantly enhancing overall system performance. During EKF-based tracking, the system consistently produced reliable predictions even under conditions involving directional changes, acceleration, or temporary target loss. In conclusion, the integration of YOLOv8n-based object detection with EKF-supported tracking has resulted in an effective solution that optimizes system accuracy, stability, and real-time operability. This study enhances the autonomous mission capabilities of UAVs and presents an innovative approach that surpasses conventional methods in the literature. The findings obtained lay a strong foundation for future flight tests on actual UAV platforms and demonstrate the system's potential for applications in defense, security, and reconnaissance domains.