Gun detection in surveillance videos using deep neural networks

The ongoing epidemic of gun violence worldwide has compelled various agencies, businesses and consumers to deploy closed-circuit television (CCTV) surveillance cameras in attempt to combat this epidemic. An active-based CCTV system extends this platform to autonomously detect potential firearms within a video surveillance perspective. However, autonomously detecting a firearm across varying CCTV camera angles, depth and illumination represents an arduous task which has seen limited success using existing deep neural networks models. This challenge is in part due to the lack of available contextual hand gun information from CCTV images, which remains unresolved. As such, this paper introduces a novel large scale dataset of hand guns which were captured using a CCTV camera. This dataset serves to substantially improve the state-of-the-art in representation learning of hand guns within a surveillance perspective. The proposed dataset consist of 250 recorded CCTV videos with a total of 5500 images. Each annotated CCTV image realistically captures the presence of a hand gun under 1) varying outdoor and indoor conditions, and 2) different resolutions representing variable scales and depth of a gun relative to a cameras sensor. The proposed dataset is used to train a single-stage object detector using a multi-level feature pyramid network (i.e. M2Det). The trained network is then validated using images from the UCF crime video dataset which contains real-world gun violence. Experimental results indicate that the proposed dataset increases the average precision of gun detection at different scales by as much as 18% when compared to existing approaches in firearms detection.