Violence detection in videos for an intelligent surveillance system using MoBSIFT and movement filtering algorithm

• Published in: Pattern analysis and applications : PAA Vol. 23; no. 2; pp. 611 - 623
• Main Authors: Febin, I. P., Jayasree, K., Joy, Preetha Theresa
• Format: Journal Article
• Language: English
• Published: London Springer London 01.05.2020; Springer Nature B.V
• Subjects: Algorithms; Complexity; Computer Science; Computer vision; Feature extraction; Filtration; Histograms; Motion perception; Movement; Optical flow (image analysis); Pattern Recognition; Security; Surveillance; Theoretical Advances; Violence; Action recognition; Abnormal activity detection; Video content analysis; Video event detection; Violence detection
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-019-00821-3

Action recognition is an active research area in computer vision as it has enormous applications in today’s world, out of which, recognizing violent action is of great importance since it is closely related to our safety and security. An intelligent surveillance system is the idea of automatically recognizing suspicious activities in surveillance videos and thereby supporting security personals to take up right action on the right time. Under this area, most of the researchers were focused on people detection and tracking, loitering, etc., whereas detecting violent actions or fights is comparatively a less studied area. Previous works considered the local spatiotemporal feature extractors; however, it accompanies the overhead of complex optical flow estimation. Even though the temporal derivative is a fast alternative to optical flow, it alone gives very low accuracy and scales-dependent result. Hence, here we propose a cascaded method of violence detection based on motion boundary SIFT (MoBSIFT) and movement filtering. In this method, the surveillance videos are checked through a movement filtering algorithm based on temporal derivative and avoid most of the nonviolent actions from going through feature extraction. Only the filtered frames may allow going through feature extraction. In addition to scale-invariant feature transform (SIFT) and histogram of optical flow feature, motion boundary histogram is also extracted and combined to form MoBSIFT descriptor. The experimental results show that the proposed MoBSIFT outperforms the existing methods in accuracy by its high tolerance to camera movements. Time complexity has also proved to be reduced by the use of movement filtering along with MoBSIFT.