Abstract
Real violence is unwanted content in video portals as it isforensically relevant in video surveillance systems. Naturally,both domains have to deal with mass data which makes thedetection of violence by hand an impossible task. We introduceone component of a system for automated violence detectionfrom video content: the differentiation of real violenceand martial arts videos. In particular, we introduce two newfeature transformations for jitter detection and local interestpoint detection with Gestalt laws. Descriptions are classifiedin a two-step machine learning process. The experimentalresults are highly encouraging: the novel features performexceptionally well and the classification process practicallyacceptable recall and precision.
Reference
Hörhan, M., & Eidenberger, H. (2013). New Content-Based Features for the Distinction of Violent Videos and Martial Arts. In Proceedings of the International Conference on Acoustics, Speech, and Signal Processing. IEEE, Austria. IEEE Press. http://hdl.handle.net/20.500.12708/54590