PhD forum: Investigating the performance of a multi-modal approach to unusual event detection

Abstract

In this paper, we investigate the parameters underpinning our previously presented system for detecting unusual events in surveillance applications [1]. The system identifies anomalous events using an unsupervised data-driven approach. During a training period, typical activities within a surveilled environment are modeled using multi-modal sensor readings. Significant deviations from the established model of regular activity can then be flagged as anomalous at run-time. Using this approach, the system can be deployed and automatically adapt for use in any environment without any manual adjustment. Experiments carried out on two days of audio-visual data were performed and evaluated using a manually annotated ground-truth. We investigate sensor fusion and quantitatively evaluate the performance gains over single modality models. We also investigate different formulations of our cluster-based model of usual scenes as well as the impact of dynamic thresholding on identifying anomalous events. Experimental results are promising, even when modeling is performed using very simple audio and visual features.