Weapon Detection with Deep Learning and Computer Graphics.

Abstract

In the last decade, deep learning and its application in computer vision have shown a big increase in popularity, due to the high performance displayed in multiple experiments and elds of application, even reaching better results than classic computer vision methods in some of them. One of these elds of application is weapon detection and surveillance, where common metal detectors are broadly used, but they need an operator to monitor their results, as well as having the problem of detecting every metallic object. For these reasons, some automatic weapon detectors have started to appear, reaching good results. This project is focused on developing an automatic weapon detector using deep learning, and at the same time, addressing some of the main problems detectors in the state of the art have encountered: the lack of labelled images for training these detectors, as well as having a big rate of false positives when using them with images from di erent perspectives. The solutions proposed for these problems are based on the use of synthetic data and anomaly detection techniques. A complete work ow has been developed to create synthetic image datasets using a graphic engine, capable of generating synthetic data of arbitrary size by developing only a virtual environment. Also, some anomaly detection techniques have been proposed to lter the amount of false detections from the weapon detector. A whole system has been developed, consisting of a detector and the aforementioned anomaly detection techniques, capable of detecting weapons in several camera perspectives. As a result, the system reports high precision in the same perspective it was trained with. When experimenting with di erent perspectives, the detector shows a big rate of false detections, but the proposed anomaly detection techniques are able to reduce this rate to less than a third.