Accurate gunshot detection in urban environments using blind deconvolution

Acoustic gunshot detection systems when deployed in urban environments need to account for reverberations in the acoustic channels. In order to improve the detection performance in such environments, we investigate the potential deployment of a two-stage methodology comprising of a blind system identification and deconvolution (BSID) stage followed by a support vector machine (SVM)-based classification stage. Given that gunshot detection systems are typically equipped with a microphone array, BSID module makes use of the spatial diversity achieved through multiple recordings of the same acoustic event and implements a maximum likelihood approach to the estimation and subsequent mitigation of the reverberation effects. In order to simulate gunshots' audio recordings in an urban environment, we use audio clips of gunshots with negligible reverberations that were recorded in an open area, and pass them through experimentally measured channel responses corresponding to a semi-enclosed outdoor courtyard. Using an extensive database of gunshots as well as outsider signals, our results indicate that introducing the BSID module achieves a dramatic improvement in detection performance, albeit under constraints necessary for working of the BSID approach. We achieve an impressive detection performance of more than 99% and a negligible false alarm rate even in the face of strong reverberations.