Automated recognition of firearms in surveillance video

CCTV surveillance systems are being deployed in workplaces, urban areas and almost every public space. The number of CCTV video streams surpasses the ability of a human operator to watch and analyse the situation carefully with respect to potentially dangerous situations, such as “Active Shooter Events”. Such events as the tragedy in the movie theatre in Colorado (USA) or Oslo (Norway) require an immediate response. In this paper, we propose and benchmark an algorithm that is capable of detecting a person carrying an uncovered firearm and alerting the CCTV operator of a potentially dangerous situation. We present the limitations and difficulties for such an image analysis application, discuss the construction of the proposed algorithm, and show the numerical results in terms of sensitivity and specificity.