Incremental behavior modeling and suspicious activity detection

We propose and evaluate an efficient method for automatic identification of suspicious behavior in video surveillance data that incrementally learns scene-specific statistical models of human behavior without requiring storage of large databases of training data. The approach begins by building an initial set of models explaining the behaviors occurring in a small bootstrap dataset. The bootstrap procedure partitions the bootstrap set into clusters then assigns new observation sequences to clusters based on the statistical tests of HMM log likelihood scores. Cluster-specific likelihood thresholds are learned rather than set arbitrarily. After bootstrapping, each new sequence is used to incrementally update the sufficient statistics of the HMM it is assigned to. In an evaluation on a real-world testbed video surveillance dataset, we find that within 1 week of observation, the incremental methodʹs false alarm rate drops below that of a batch method on the same data. The incremental method obtains a false alarm rate of 2.2% at a 91% hit rate. The method is thus a practical and effective solution to the problem of inducing scene-specific statistical models useful for bringing suspicious behavior to the attention of human security personnel.