Among the threats that can happen in public space, the apparition of a person with a gun or a rifle in hand is a high level danger. This type of situation, hopefully very rare, provokes an important confusion, and one can observe that the transmission of the relevant information to the site security central point concerned (a train station, an airport, a police central, the security center of a large shop or a public event...) is delayed, confused, and does not allow to determine precisely the circonstumces of the incident. It does not allow LEAs  to call quickly and efficiently the appropriate Forces. Delays of 20, 30 minutes and even more are observed, delays that could have helped Forces to come earlier, to limit the incident extension, and reduce casualties.

For these reasons, we have launched by end 2018, and led these two last years, the collaborative DRAAF R&D project, with the best public research units (INRIA, ENS-ULM, CNRS, GREYC) in order to propose a solution allowing to detect very quickly these situations, from video feeds, and warn immediatly the relevant video security office.

Difficulties to realize such an application are many, due for example to the low image resolution, the moves in the video, and the partial non-visibility of the weapon, when hold in hand.

Last, in order to keep the availability of the surveillance agents, as well as their capacity to react, it is compulsory to avoid to warn them regularly and inappropriately. Hence, the project has developed approaches targeting the reduction to the minimum of the false alarms, using simultaneously different AI innovative techniques for analyzing gestures, small object detection, and body attitudes, with a strong involvement in performances of the algorithm until it can process in real time a sufficent number of cameras on a given price-effective computing unit.

Tests have been realized as shown here (people faces are hidden) :