Change Point Geometry for Change Detection in Surveillance Video

Brandon A. Mayer, Joseph L. Mundy

Abstract: A change detection algorithm is proposed based on geometric descriptors of space-time appearance discontinuities in fixed camera video. At each pixel in a video frame, intensity subsequences with similar appearance are segmented using a Hidden Semi-Markov Model (HSMM). The start of each per-pixel homogeneous subsequence, referred to as change point vertices, are then clustered across pixel locations using an efficient graph based segmentation algorithm to construct a change point hull. The geometry of the change point hull provides a discriminating feature for distinguishing coherent movement from random or stochastic appearance changes and is simultaneously a rich descriptor for reasoning about object velocity and direction. State of the art results are shown in change detection, a fundamental computer vision problem for identifying regions of video that exhibit meaningful variations as defined by the application context.