Detecting Moving Objects from Moving Background by Optical Flow Decomposition

Detecting moving objects from image sequences collected by a moving camera, e.g., onboard an unmanned aerial vehicle (UAV), is an important yet challenging problem. Existing methods based on supervised learning fall short when the labeled data are limited. To overcome such limitations, this paper proposes an unsupervised learning method based on a tensor decomposition approach. The optical flow estimated from the apparent motion of pixels between consecutive frames is decomposed into a superposition of a background, a foreground, and noise, each of which is regularized by considering their motion pattern. An ADMM-based algorithm is developed to optimally estimate these three components. The advantages of the proposed method are demonstrated by a real-world case study.