Cooperative Localisation on Android Devices by Utilising only Environmental Sound

This thesis focusses on research that gears towards Cooperative Localisation utilising only ambient sound. Sound signals can be used for Time Difference Of Arrival (TDOA) based Cooperative Localisation on mobile devices. Android has a large penetration in the worldwide market, thus developing an application that is able to run on Android devices is very interesting. Combining these arguments leads to the main research question: When utilising only environmental sound that originates from unknown positions, what techniques for Cooperative Localisation can be used on Android devices that can achieve accuracy within several metres, and what factors will influence this accuracy? In order to answer this question this thesis introduces the Cooperative Localisation on Android with ambient Sound Sources (CLASS) Algorithm. This Algorithm produces a location set for all devices and a set of directions towards the origins of the sound-events. A Histogram Based Outlier detection Algorithm is implemented to find outliers in the localisation results. The CLASS Algorithm deals with inaccurate measurement data by finding and averaging TDOA values, and localisation results, that are inliers. To our best knowledge no prior work has utilised Android for Cooperative Localisation. The following question is therefore posed: What are the technical limitations of utilising a non Real-Time Operating System like Android for Cooperative Localisation that achieves accuracy within several metres? This thesis argues that input latency and poor time synchronisation are the main limitation of the Android Operating System (OS) for it’s use in Cooperative Localisation by sound. Input latency in Android suffers from large jitters that cannot be predicted and corrected. The following technical limitations of Android, ordered from most to least significant, contribute to TDOA measurement inaccuracies: (i) audio input latency, (ii) poor time synchronisation, (iii) difference in microphone gain per device, (iv) delays in recording time-stamps, (v) implementing a Digital Signal Processor (DSP) like Fast Fourier Transformation (FFT), (vi) noise in the form of peaks in the microphone signal. These limitations can result in erroneous TDOA measurements and are dealt with in the CLASS Algorithm and Android application. The accuracy of the CLASS Algorithm was assessed with an outdoor experiment. Sound signals were generated with an air horn at twenty locations around a constellation of 16 Nexus-7 tablets. Four sound-events were generated at each location. The sound-events were distributed along a circle with a radius of 46 m The devices were placed in a 12 m× 12 m grid with a perpendicular inter-device distance of 4 m. A mean Root Mean Square Error (RMSE) of 2.4 m with a standard deviation of 0.21 m is achieved. The mean RMSE of the estimated directions is 76.24 ◦ with a standard deviation of 1.28 ◦. These results can be improved in future work by elaborating on different parts in the CLASS Algorithm and Android application. Thesis Supervisor: Paul Havinga Title: Full Professor