Bayesian detection of acoustic muzzle blasts

Abstract

Acoustic detection of gunshots has many security and military applications. Most gunfire produces both an acoustic muzzle-blast signal as well as a high-frequency shockwave. However some guns do not propel bullets with the speed required to cause shockwaves, and the use of a silencer can significantly reduce the energy of muzzle blasts; thus, although most existing commercial and military gunshot detection systems are based on shockwave detection, reliable detection across a wide range of applications requires the development of techniques which incorporate both muzzle-blast and shockwave phenomenologies. The detection of muzzle blasts is often difficult due to the presence of non-stationary background signals. Previous approaches to muzzle blast detection have applied pattern recognition techniques without specifically considering the non-stationary nature of the background signals and thus these techniques may perform poorly under realistic operating conditions. This research focuses on time domain mo eling of the non-stationary background using Bayesian auto-regressive models. Bayesian parameter estimation can provide a principled approach to non-stationary modeling while also eliminating the stability concerns associated with standard adaptive procedures. Our proposed approach is tested on a synthetic dataset derived from recordings of actual background signals and a database of isolated gunfire. Detection results are compared to a standard adaptive approach, the least-mean squares (LMS) algorithm, across several signal to background ratios in both indoor and outdoor conditions. © 2009 SPIE.

DOI
10.1117/12.818547
Year