Neural network based threat assessment for automated visual surveillance

In automated visual surveillance systems (AVSS), reliable detection of suspicious human behavior is of great practical importance. Many conventional classifiers have shown to perform inadequately because of unpredictable nature of human behavior. Flexible models such as artificial neural network (ANN) models can perform better; however, computational requirement of ANN models can be prohibitively large for realtime video processing. It is interesting to construct a small-sized ANN classifier that can perform well for threat assessment in video-based surveillance system. In this paper, modified probabilistic neural network (MPNN) is introduced that can achieve reliable classification, with significantly reduced computation. Experiment on visual surveillance application shows that MPNN achieves good classification but with much reduced computation compared to other ANN models. In this application, trajectory profile and motion history image information from the observed human subject are used for threat assessment.