Real-time implementation of moving object detection in UAV videos using GPUs

Unmanned aerial vehicles (UAVs) are being increasingly used for video surveillance and remote sensing. Moving object detection is an important algorithm for many such applications. Real-time processing of moving object detection is required for various decision-making tasks in many of these applications. However, being compute-intensive in nature, it is difficult to process high-resolution UAV-sourced videos in real-time. GPU vendors regularly release newer architectures with new features to speed up various kinds of applications. Hence, it becomes imperative to explore parallel implementations of such algorithms using the new GPU architectures. This paper describes parallel implementation strategies for algorithms like feature detection, feature matching, image transformation, frame differencing, morphological processing and connected component labeling which are used to detect moving objects in UAV-sourced videos. The implementation is tested on different NVIDIA GPU microarchitectures (Fermi, Maxwell, and Pascal). Experimental results show the achieved frame processing rates of 43.1 fps, 35.5 fps and 9.1 fps for 1080p videos on Pascal, Maxwell, and Fermi microarchitectures respectively.