Deep multi-level feature pyramids: Application for non-canonical firearm detection in video surveillance

The epidemic of gun violence worldwide necessitates the need for an active-based video surveillance network to combat this crime. In this context, autonomously detecting handguns is crucial in capturing firearm-related crimes. However, current object detectors using deep learning are unable to capture handguns at different scales in an unconstrained environment. Hence, this paper puts forward an enhanced deep multi-level feature pyramid network that addresses the difficulty in inferring handguns from a non-canonical perspective. We first construct a dataset containing handguns in an unconstrained environment for representation learning. The dataset is constructed from a set of 250 recorded videos and with over 2500 distinct labeled frames. Crucially, these labeled frames account for the absence of a proper video surveillance-based handgun dataset. We then train the dataset on a multi-level multi-scale object detector, i.e., M2Det. We further improve the performance of M2Det by: (1) Enhancing the base features by concatenating shallow, medium and deep features from the backbone according to its relative receptive field; (2) Implementing generalized intersection-over-union as its localization loss; and (3) Integrating Focal Loss as its classification loss to improve detection of small-scale handguns. Experiments on a challenging video surveillance test dataset demonstrate that the proposed model achieves 87.42% accuracy. In addition, we implement adaptive surveillance image partitioning to redetect handguns at specific regions. This method potentially solves the challenge of sporadically poor real-world handgun classifications. This model is capable of pioneering non-canonical handgun detection for active-based video surveillance systems. The dataset and trained models are available at:https://github.com/MarcusLimJunYi/Monash-Guns-Dataset.