Abnormal behavior detection using a multi-modal stochastic learning approach

This paper presents a new approach to trajectory-based abnormal behavior detection (ABD). While existing techniques include position in the feature vector, we propose to estimate the probability distribution locally at each position, hence reducing the dimensionality of the feature vector. Local information derived from accumulated knowledge for a particular position is integrated in the distribution enabling context-based decision for ABD. A stochastic competitive learning algorithm is employed to estimate the local distributions of the feature vector and the location of the distribution modes. The proposed algorithm is tested on the detection of driving under the influence of alcohol. The performance of the new algorithm is evaluated on synthetic data. First the local stochastic learning algorithm is compared to its global variant. Then it is compared to the Kohonen self organizing feature maps. In both cases, the proposed algorithm achieves higher detection rates (at the same false alarm rate) with fewer clusters.