Motion-Segmentation-Based Change Detection

Detecting regions of change in images of the same scene taken at different times is of widespread interest. Important applications of change detection include video surveillance, remote sensing, medical diagnosis and treatment. Change detection usually involves image registration, which is aimed at removing meaningless changes caused by camera motion. Image registration is a hard problem due to the absence of knowledge about camera motion and objects in the scene. To address this problem, this paper proposes a novel motion-segmentation based approach to change detection, which represents a paradigm shift. Different from the existing methods, our approach does not even need image registration since our method is able to separate global motion (camera motion) from local motion, where local motion corresponds to regions of change while regions with only global motion will be classified as ‘no change’. Hence, our approach has the advantage of robustness against camera motion. Separating global motion from local motion is particularly challenging due to lack of prior knowledge about camera motion and the objects in the scene. To tackle this, we introduce a motion-segmentation approach based on minimization of the coding length. The key idea of our approach is as below. We first estimate the motion field by solving the optical flow equation; then we segment the motion field into regions with different motion, based on the minimum coding length criterion; after motion segmentation, we estimate the global motion and local motion; finally, our algorithm outputs regions of change, which correspond to local motion. Experimental results demonstrate the effectiveness of our scheme.