Directional Filtering for Sequential Image Analysis

Image sequences produced by, e.g., electro-optical sensors are being used increasingly for continuous quality monitoring and surveillance. This letter considers the important problem of multiframe change point detection of signals with unknown dynamics. We propose a new CUSUM-based (cumulative SUM) detector for detecting noise-obscured signals in a sequence of frames. Our sequential detector adopts directional filtering, a search philosophy of interesting pixels, of state-of-the-art detector that employs dynamic programming (DP) for so-called track-before-detect (TBD) processing. In case of a persistent change, a detailed comparison reveals our CUSUM-based detector superior to the DP-based TBD detector.