Unsupervised anomalous behavior detection using spatial-temporal interest points

We presented a novel framework for automatic behavior profiling and unsupervised anomaly detection in a large video set using spatial-temporal interest points. The framework consisted of the following key components: 1) A compact and effective behavior representation method was developed based on spatial-temporal interest points. 2) The natural grouping of behavior patterns were discovered through a novel hierarchical topic model, Latent Dirichlet Co-Clustering (LDCC), which were used to connect three elements in visual surveillance: spatial-temporal interest points, simple atomic actions and behavior patterns. Atomic actions were modeled as distributions over spatialtemporal interest points, and behavior patterns as distributions over atomic actions. 3) An online anomaly measure was introduced to detect abnormal behavior, whereas normal behavior patterns were recognized by runtime accumulative visual evidence using Likelihood Ratio Test (LRT) method. The experimental results demonstrated the effectiveness and robustness of our approach using noisy and sparse datasets collected from outdoor surveillance scenarios. © 2011 ISSN 1881-803X.