A hardware based gunshot sound detection system

Abstract

With the development of semiconductor technology, embedded systems' capacity of operations also increases day by day. In this way, small-sized devices are able to perform complex works. As a result, people take advantage of embedded systems in a wide variety of areas to enhance the life quality of living. As a result of technological developments, the use of tools that assist law enforcement officers in crime detection is also increasing. Available gunshot detection systems mainly focuses on preventing illegal hunting, decreasing crime rates in public space, and detecting gunshot direction in battlefield areas. When the literature is examined, it is seen that machine learning methods are used in the studies used in gunshot sound detection. However, the number of hardware-based systems used in gunshot sound detection is quite a few and mostly simple methods such as cross-correlation threshold, edge detection are implemented. In this work, it is aimed to realize a gunshot sound detection system on hardware. In this context, it is aimed to select the system that uses the advantages of machine learning methods and is the most suitable for implementation on the hardware. When the literature is examined, it has been observed that the mel coefficients, signal energy and zero crossing properties perform well in determining the gunshot sound. For this reason, the mentioned features were obtained from the audio signal and used in k-nearest neighbors (k-NN) and support vector machines (SVM) classification algorithms. An accuracy rate of 96.1538% was obtained with the k-NN classifier and 91.3462% with the SVM classifier.