Robust Online Change-point Detection in Video Sequences

We present the Cumulative Sum (CUSUM) stopping rule, applied to Computer Vision problems, to automatically detect changes in either parametric or nonparametric distributions, online or off-line. Our approach is based on using the previously received data of the sequence to detect a change in data that are to be received. We assume that no significant change has occurred up to an unknown time instance. Then a change in the distribution of the observations occurs and the objective is to estimate this instance. We test the hypotheses of no change occurs vs. a change occurs at the current frame, which is done by the CUSUM stopping rule. We apply our framework to the case of continuous 3D hand tracking, where the high dofs, the fast finger articulations, the large rotations and the frequent occlusions often cause error accumulation. Also we illustrate the performance of our approach in video segmentation, and specifically in segmentation of fingerspelling in American Sign Language (ASL) videos.