The 22nd International Conference on Information Fusion, held between the 2nd and 5th July in Ottawa, Canada, brought together some of the best minds in the sensor, data, information and knowledge fusion domains. Contributing to this event, Crime and Security Research (CSRI) PhD Student Marc Roig Vilamala presented his paper “A Pilot Study on Detecting Violence in Videos Fusing Proxy Models”.
Marc reported results from research that applies state-of-the-art machine learning and reasoning to a dataset of crime (CCTV) videos in order to detect violent events. This novel pilot study has begun initial exploration into creating a system in which anomalies in CCTV feeds such as fighting, road accidents and other unusual events in the street are identified.
The presentation noted the difficulties of such a project in responding to the wide spectrum of data caused by variables in the activity, environments and times of day that the feeds recorded.
Responding to these challenges, the researchers used two off-the-shelf neural networks, rather than developing their own training. The first neural network detected activities from fragments of video of around half a second, whilst the second identified people and other objects within each frame. To detect when an anomaly was happening, the researchers applied manually defined generic rules to fuse both of these outputs.