Multi-Observation Continuous Density Hidden Markov Models for Anomaly Detection in Full Motion Video.

摘要

An increase in sensors on the battle eld produces an abundance of collected data that overwhelms the processing capability of the DoD. Automated Visual Surveillance (AVS) seeks to use machines to better exploit increased sensor data, such as by highlighting anomalies. In this thesis, we apply AVS to overhead Full Motion Video (FMV). We seek to automate the classi cation of soldiers in a simulated combat scenario into their agent types. To this end, we use Multi- Dimensional Continuous Density Hidden Markov Models (MOCDHMMs), a form of HMM which models a training dataset more precisely than simple HMMs. MOCDHMMs are theoretically developed but thinly applied in literature. We discover and correct three errors which occur in HMM algorithms when applied to MOCDHMMs but not when applied to simple HMMs. We o er three xes to the errors and show analytically why they work. To show the xes e ective, we conduct experiments on three datasets: two pilot experiment datasets and a simulated combat scenario dataset. The modi ed MOCDHMM algorithm gives statistically signi cant improvement over the standard MOCDHMM: 5% improvement in accuracy for the pilot datasets and 3% for the combat scenario dataset. In addition, results suggest that increasing the number of hidden states in an MOCDHMM classi er increases the separability of the classes but also increases classi er bias. Furthermore, we nd that classi cation based on tracked position alone is possible and that MOCDHMM classi ers are highly resistant to noise in their training data.