Detection and Localisation of Gunshots Using Sound Data

Sammanfattning: The goal of this master's thesis is to detect and position sharp sounds using Axis speakers with built-in microphones. Sharp sounds of special interest are gunshots. The system needs at least five speakers to function and is designed for usage in indoor environments. The project follows a pipeline in order to position sound sources containing recording, synchronisation, detection of gunshot in sound data, and positioning of the sound. Detection of gunshots in recorded files is done by a binary classification with a deep neural network created in Python. The algorithms for positioning are implemented in MATLAB. The final neural network has an accuracy of 98%. It is pretrained by VGG-team with data from ILSVR and transfer learning is applied to fit the model for gunshot data. After testing a few methods to synchronise the speakers and to calculate the position of the sound source, the final system has a mean error of 0.28 m. The model's precision is adequate for large areas.