Duration Dependent Codebooks for Change Detection

This paper describes a supervised system for pixel-level change detection for fixed, monocular surveillance cameras. Per-pixel intensity sequences are modeled by a class of Hidden Semi-Markov Models, Duration Dependent Hidden Markov Models (DDHMMs), to accurately account for stochastically periodic phenomena prevalent in real-world video. The per-pixel DDHMMs are used to assign discrete state labels to pixel intensity sequences which summarize the appearance and temporal statistics of the observations. State assignments are then used as a features for constructing per-pixel code books during a training phase to identify changes of interest in new video. The per-pixel intensity model is validated by showing superior predictive performance to pixel representations commonly used in change detection applications. A new data set is presented which contain dynamic, periodic backgrounds with larger time scale variability than previous data sets and the proposed method is compared to state-of-theart change detection methods using the new videos.