Machine learning inspired efficient acoustic gunshot detection and localization system

Abstract

Gunshot detection and localization is a frontier technology in security systems. With an increasing rate of shootings globally, gunshot events and directional awareness are crucial for the law enforcement agencies for a timely response. This paper presents a real-time computational efficient gunshot detection and localization system. First, the performance of Mel-frequency cepstral coefficients, linear prediction coefficients, Gammatone cepstral coefficients, and spectral centroid as an audio feature for acoustic gunshot detection is thoroughly analyzed. Then, a bagged tree ensemble and support vector machine classifiers are trained and tested on a diverse gunshot database under different SNR settings, using a 10-fold validation technique. The detection accuracy of 97.3% with a sensitivity of 0.978 and a specificity of 0.988 is achieved. The test-train curves corroborate the fitness and generalization of the trained detection model. After the detection, the localization is performed by calculating the arrival time difference using a general cross-correlation phase transform. Finally, the system is implemented on an experimental test-bed for real-time performance evaluation. Field tests indicate the proposed system's effectiveness to detect and localize a gunshot in 0.7-1 seconds.