PROJECT TITLE :

Video Behavior Profiling for Anomaly Detection

ABSTRACT:

The goal of this study is to solve the challenge of modeling video behavior acquired in surveillance cameras for online normal behavior recognition and anomaly detection applications. Without any manual labeling of the training data set, a novel framework for automatic behavior profiling and online anomaly sampling/detection is established. The framework is made up of the following essential elements: 1) Using discrete-scene event detection, a compact and effective behavior representation approach is devised. The similarity of behavior patterns is assessed using a Dynamic Bayesian Network to model each pattern (DBN). 2) A unique spectral clustering approach with unsupervised model selection and feature selection on the eigenvectors of a normalized affinity matrix is used to uncover the natural grouping of behavior patterns. 3) A composite generative behavior model is built that can generalize from a short training set to account for changes in normal behavior patterns that aren't visible. 4) To detect abnormal behavior, a runtime accumulative anomaly measure is introduced, but normal behavior patterns are detected when adequate visual evidence is present using an online Likelihood Ratio Test (LRT) technique. This ensures that anomalies are detected and typical behavior is recognized in the shortest time feasible. Experiments with noisy and sparse data sets obtained from both interior and outdoor monitoring scenarios illustrate the usefulness and robustness of our approach. In particular, it is demonstrated that in detecting abnormality from an unseen video, a behavior model trained using an unlabeled data set outperforms those trained using the same but labeled data set. Our online LRT-based behavior identification strategy outperforms the commonly used Maximum Likelihood (ML) method in distinguishing ambiguity among different online behavior classes, according to the results.