A multi-temporal framework for high-level activity analysis: Violent event detection in visual surveillance

This paper presents a novel framework for high-level activity analysis based on late fusion using multi-independent temporal perception layers. The method allows us to handle temporal diversity of high-level activities. The framework consists of multi-temporal analysis, multi-temporal perception layers, and late fusion. We build two types of perception layers based on situation graph trees (SGT) and support vector machines (SVMs). The results obtained from the multi-temporal perception layers are fused into an activity score through a step of late fusion. To verify this approach, we apply the framework to violent events detection in visual surveillance and experiments are conducted by using three datasets: BEHAVE, NUS-HGA and some videos from YouTube that show real situations. We also compare the proposed framework with existing single-temporal frameworks. The experiments produced results with accuracy of 0.783 (SGT-based, BEHAVE), 0.702 (SVM-based, BEHAVE), 0.872 (SGT-based, NUS-HGA), and 0.699 (SGT-based, YouTube), thereby showing that using our multi-temporal approach has advantages over single-temporal methods.