English SIANI English English

Design and gradual implantation methodology

for a violence detector in video


Start Date: 1/1/2012 Funding: Spanish Ministry for Economy and Competitiveness / European Regional Development Fund
End Date: 31/12/2014 Project Code: TIN2011-24367


Oscar Deniz Suarez
Ismael Serrano Gracia
Gloria Bueno Garcia
Jesus Salido
Noelia Vallez Enano
Fco. Mario Hernandez Tejera

External researchers:
Rahul Sukthankar

Collaborating institutions:

Infaimon S.L.
Cuerpo Nacional de Policía
Axis Communications


Project description:

In the last years, efforts in computer vision have allowed researchers to have robust and versatile detectors that can be now used as components in many different applications. Two well-known examples are face and person detectors, which are available in open-source computer vision libraries. On the other hand, the increasing rate of video content generation and the availability of more computing power have led to the development of action recognition in video. Most of the previous work on action recognition focuses on simple human actions like walking, jumping or hand waving. Despite its potential usefulness, violent action detection has been less studied.

This project aims at developing a violence detector that is as robust and versatile as available face and person detectors. A violence detector has immediate applicability both in the surveillance domain (prisons, eldercare facilities, sport fields, conflictive underground stations and even for mobile cameras that may be sold to teenagers) and for rating/tagging online video content.

This project also aims at adopting a new implementation methodology which we consider more appropriate for videosurveillance applications. Instead of placing more importance on detection capacity, the methodology is based on controlling false alarms, which is what ultimately determines acceptability of the detection system. To this end, we propose a novel method for generating a cascade of classifiers based upon false positives gathered in the specific scenario where the detector is deployed.



  • "Fast Fight Detection" accepted in PLoS ONE
  • "Fast Violence Detection in Video", accepted in VISAPP 2014
  • "False positive reduction in detector implantation", accepted in AIME 2013
  • "VISILAB at MediaEval 2013: Fight detection", accepted in MediaEval 2013
    Multimedia Benchmark Workshop
  • Our Hockey fight video dataset is already being used by other researchers for comparing violence/fight detectors! See Violent Flows: Real-Time Detection of Violent Crowd Behavior.
  • The Principal Researcher has been invited to an AAAI Spring Symposium 2013 (Stanford, March 25-27, 2013) to give a talk about Fight Detection


Ver en Español