Target detection and tracking from unmanned aerial vehicle cameras using embedded GPU

In surveillance, camera aerial vehicles are a common tool. The majority of these surveillance systems follow things in two phases: they identify and detect their targets first, and then they follow them in the live video feed. However, modern object identification methods often use deep learning models trained on massive picture datasets and operate on GPU powered platforms with tremendous processing capacity. Additionally, managing feature matching and association in object tracking introduces additional activities into the system, hence impacting the system's real-time performance. For wide area surveillance (WAS) applications, vision-based object recognition and target tracking is crucial for identifying and following things on ground targets. But since these unmanned aerial vehicles (UAVs) fly so far above the ground, they distort the size of things on the ground. Consequently, it is necessary to have a sensitive object detection detector that can identify these minute ground objects and thoroughly scan characteristics. Furthermore, since CNN based object detectors use deep learning and need a lot of computing and mathematical modeling, they might be difficult to use on embedded devices. The before mentioned issues that may impact object identification and tracking system performance were thoroughly investigated in this thesis. The absence of extensive UAV datasets including object size, brightness, and diversity in form and texture details in the photos captured from airplanes is one of the issues with target identification. Accuracy in object tracking is impacted by performance problems brought on by deep learning approaches' high computational complexity. This thesis uses the RNN model for target tracking and the YOLOV4 target detection model in embedded devices like the NVIDIA JETSON AGX XAVIER to design an efficient, quick, daytime, and real-time target detection and tracking system that can operate quickly and efficiently with the model optimization approach in embedded systems.