Topic models such as Latent Dirichlet Allocation (LDA) are used extensively for modelling multi-object behaviour and anomaly detection in busy scenes. However, existing topic models suffer from the sensitivity problem, where they are unable to detect anomalies that are mixed in with large numbers of co-occurring normal behaviours. Also at issue is the localisation problem, where anomalies are detected but not localised within a given video clip. To address these two problems this paper proposes a novel region LDA model, which encodes the spatial awareness that is ignored by conventional topic models. Both scene decomposition and behavioural modelling are simultaneously performed. Consequentially, abnormality is detected per-region rather than for the entire scene, resolving both the sensitivity and localisation issues. Experiments conducted on busy real world scenes demonstrate the superiority of the proposed model.
|Number of pages||6|
|Publication status||Published - 2010|
|Event||ACM international workshop on Multimodal pervasive video analysis - Firenze, Italy|
Duration: 25 Oct 2010 → 29 Oct 2010
Conference number: 1
|Workshop||ACM international workshop on Multimodal pervasive video analysis|
|Period||25/10/10 → 29/10/10|