Kan Ouivirach, Ph.D.

We propose and experimentally evaluate a new method for detecting shadows using a simple maximum likelihood formulation based on color information. We first estimate, offline, a joint probability distribution over the difference in the HSV color space between pixels in the current frame and the corresponding pixels in a background model, conditional on whether the pixel is an object pixel or a shadow pixel. Given the learned distribution, at run time, we use the maximum likelihood principle to… Show more

We propose and experimentally evaluate a new method for detecting shadows using a simple maximum likelihood formulation based on color information. We first estimate, offline, a joint probability distribution over the difference in the HSV color space between pixels in the current frame and the corresponding pixels in a background model, conditional on whether the pixel is an object pixel or a shadow pixel. Given the learned distribution, at run time, we use the maximum likelihood principle to classify each foreground pixel as either shadow or object. In an experimental evaluation, we find that the method outperforms standard methods on three different real-world video surveillance data sets. We conclude that the proposed shadow detection method would be an extremely effective component in an intelligent video surveillance system. Show less

A new method for automatic identification of suspicious behavior in video surveillance data. The approach works by constructing scene-specific statistical models explaining the behaviors occurring in a small bootstrap data set. It partitions the bootstrap set into clusters then assigns new observation sequences to clusters based on statistical tests of HMM log likelihood scores.

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 data set. The bootstrap procedure partitions the bootstrap set into clusters then assigns new observation sequences to… Show more

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 data set. The bootstrap procedure partitions the bootstrap set into clusters then assigns new observation sequences to clusters based on 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 data set, we find that within one week of observation, the incremental method's 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. Show less

A new method for clustering human behaviors that is suitable for bootstrapping an anomaly detection module for intelligent video surveillance systems. The method uses dynamic time warping, agglomerative hierarchical clustering, and hidden Markov models to provide an initial partitioning of a set of observation sequences then automatically identifies where to cut off the hierarchical clustering dendrogram.