Online Anomaly Detection over Live Social Video Streaming

• DOI: 10.48550/arxiv.2401.08615
• Publication Date: 2024-01-01
• Authors: Chengkun He; Xiangmin Zhou; Chen Wang; Iqbal Gondal; Jie Shao; Xun Yi
• Keywords: Streaming Data; Anomaly (physics); Online video; Live Streaming

Social video anomaly is an observation in streams that does not conform to a common pattern of dataset's behaviour. detection plays critical role applications from e-commerce e-learning. Traditionally, techniques are applied find anomalies broadcasting. However, they neglect the live social which contain interactive talk, speech, or lecture with audience. In this paper, we propose generic framework for effectively online detecting Anomalies Over Video LIve Streaming (AOVLIS). Specifically, novel deep neural network model called Coupling Long Short-Term Memory (CLSTM) adaptively captures history behaviours presenters and audience, their mutual interactions predict behaviour at next time point over streams. Then well integrate CLSTM decoder layer, new reconstruction error-based scoring function $RE_{IA}$ calculate score each segment detection. After that, update scheme incrementally maintains decoder. Moreover, design upper bound ADaptive Optimisation Strategy (ADOS) improving efficiency our solution. Extensive experiments conducted prove superiority AOVLIS.