Detecting stabbing by a deep learning method from surveillance videos

Abstract

Stabbing is one of culprits threatening public safety. Once it happens, it will make immeasurable consequences in a short period of time. In order to strengthen the supervision of public safety and prevent the emergence of stabbing, the tool detection technology can play a vital role. The existing methods for tool detection are metal detector and X-ray detector, which are applicable to stations, airports and other specific areas but not feasible in public areas crowds of people. This paper proposes the use of a deep learning method with high precision and speed for tool detection and by comparison finally chooses the YOLOV3 method for tool detection in public areas. To validate the performance of YOLOV3 method, a total of 1,738 images of different tools are acquired by simulating real scenes and the web crawler technology. Meanwhile, the number of samples are amplified by image enhancement techniques, and a datasets of 21,000 images are filtered. To improve tool detection accuracy, this paper proposes a method that combines hand features and tool features into new features. Experiments have shown that the detection accuracy is improved by 2.57 % with these new features.