Our pipeline, orchestrated with Python, operates with three concurrent threads: one to continuously capture audio received from an attached microphone and put two seconds worth of said audio onto an audio analysis queue; one to analyze sound samples retrieved from the audio analysis queue and verify whether or not a gunshot occurred in a given sample; and finally one to dispatch an SMS alert message to a predetermined list of phone numbers if a gunshot was detected in the segment of audio most recently analyzed.
Our short message service (SMS) pipeline for detecting gunshots was deployed on a Raspberry Pi 3 Model B+ connected to an AT&T USBConnect Lightning Quickstart SMS modem as well as a Sizheng omnidirectional USB microphone.
We trained three different models on a set of nearly 60,000 two-second audio samples to distinguish gunshots from other noises. The first model was a one-dimensional convolutional neural network that takes a two-second time sequence of audio as input. The second and third models were two-dimensional convolutional neural network that takes a spectrogram of a two-second audio sample as input. For inference in practice, the decision on a sample's class was reached by majority-rules consensus between the three models.
Our training features and samples along with our trained models can be found on Dataverse.
This repository contains the source code accompanying the paper Low Cost Gunshot Detection using Deep Learning on the Raspberry Pi.
Please cite as:
@INPROCEEDINGS{9006456,
author={A. {Morehead} and L. {Ogden} and G. {Magee} and R. {Hosler} and B. {White} and G. {Mohler}},
booktitle={2019 IEEE International Conference on Big Data (Big Data)},
title={Low Cost Gunshot Detection using Deep Learning on the Raspberry Pi},
year={2019},
pages={3038-3044},
doi={10.1109/BigData47090.2019.9006456}}