Dangerous Tool Detection for CCTV Systems
In this paper we present our work towards an effective solution for detection of dangerous objects, such as firearms or knives in a Closed Circuit Television System. We have gathered a large, manually annotated dataset of recordings supplemented by our or
- PDF / 2,341,843 Bytes
- 14 Pages / 439.37 x 666.142 pts Page_size
- 23 Downloads / 240 Views
bstract. In this paper we present our work towards an effective solution for detection of dangerous objects, such as firearms or knives in a Closed Circuit Television System. We have gathered a large, manually annotated dataset of recordings supplemented by our original artificial sample generation method. We have used this dataset for training of a convolutional neural network. We present our approach and training results. We have also implemented and present software architecture that implements the neural network. We have shown, that the convolutional neural networks are well suited even for such complex object detection task, when provided with enough training samples. Keywords: Machine learning · Convolutional neural networks Dangerous tools · Data analysis · Object detection
·
Both in the United States and in the rest of the world the crimes with use of weapons are on the rise. For the US the Mother Jones project [13] monitors Mass Shootings (or Active Shooter Incidents). The database held by the project shows significant rise of both the number of such attacks and the number of fatalities. In the EU such events are more commonly classified as terrorist attacks. And while the number of terrorist attacks is slowly declining over the last 40 years, the fraction of attacks with use of dangerous tools is on the rise. In this paper we propose an effective and fully automated system that is capable of raising of an alarm if a person holding a dangerous object (a pistol, revolver, rifle or a knife) is visible in the CCTV image. Our solution has novel features that make it interesting not only for the scientific community but also for the market use. First, we have trained our algorithms using a large and custom created dataset (which is made available for the scientific community). Second, we have consulted the requirements for the system with the end users and have focused on limiting the amount of false alarms raised, which comes at the cost of sensitivity. This assumption might be surprising, but the justification is that if a system detects only a part of events – it is still more effective than a bare This work was supported by the Polish National Center for Research and Development under the LIDER Grant (No. LIDER/354/L-6/14/NCBR/2015). c Springer Nature Switzerland AG 2020 A. Dziech et al. (Eds.): MCSS 2020, CCIS 1284, pp. 238–251, 2020. https://doi.org/10.1007/978-3-030-59000-0_18
Dangerous Tool Detection for CCTV Systems
239
CCTV solution without any automation. On the other hand – a system that raises too many false alarms is useless as it overloads the operator. Finally, in order to construct our solution we have utilized state-of-the-art techniques such as convolutional neural networks, GPU based computing and genetic algorithms making it, to our best knowledge, one of the most advanced systems of its kind. The rest of the paper is structured as follows. Section 1 provides a description of the research problem. Section 2 covers the related work in the filed. Section 3 describes our dataset and methods used for
Data Loading...
