A Hybrid Model for 3-D Gunshot Localization Using Muzzle Blast Sound Only

Abstract

Humans and animals possess the ability to roughly estimate the distance and direction of specific sound sources through their auditory organs. This study introduces an innovative hybrid model for 3-D gunshot sound localization, drawing inspiration from human sound localization models and leveraging the strengths of both microphone arrays and deep learning techniques. The proposed system employs a six-microphone array with co-centered orthogonal placement to capture the sounds of gunshot muzzle blasts. The methodology involves first estimating azimuth and elevation angles using an approximate closed-form direction-of-arrival (DoA) formula based on time-difference-of-arrival (TDoA) values obtained from the Generalized Cross Correlation with Phase Transform (GCC-Phat) algorithm. Subsequently, Mel spectrograms are derived from the captured sound data and fed into a convolutional neural network (CNN) for distance estimation. Ultimately, the direction and distance estimations are integrated to achieve the 3-D localization of the gunshot source. The proposed approach stands as a notable contribution to the literature, relying solely on the sound of a muzzle blast for localization with a single microphone array. The average distance estimation error in the range of 50-500 m is 6.87%, demonstrating an improvement compared to existing systems with an error rate of approximately 16%. This study is pioneering in its application of deep learning training on a dataset of actual explosion sounds for distance estimation in gunshot localization systems.