Online Moving Object Detection from Aerial Videos by Leveraging Geometric Constraints

The fast growing of consumer drones has prompted a lot of relevant applications such like moving object detection. Different methods have been proposed to detect moving targets in aerial videos, many rely on dense optical flow and others are based on frame differencing. They are either very time consuming or susceptible to illumination changes. In this paper, two geometric constraints are leveraged to detect moving targets for low-attitude aerial platforms. First is the homography constraint which assumes that every two consecutive frames are well related by a 3 × 3 non-singular matrix. Second is the epipolar constraint that any pixel in one image is projected onto a line in another image. Since geometric constraints widely exist in aerial videos, the proposed method is more resistant to challenges like illumination variation and background clutters. Besides, compared to multi-frame based methods, the proposed algorithm only takes a pair of consecutive frames as input, making it more efficient and easier to process fast moving targets. Extensive experiments on public data show that the proposed method not only has real-time efficiency on a single CPU, but also brings with a low false alarm rate when dealing with challenging scenarios.