Audio-based localisation for ubiquitous sensor networks

Abstract

This research presents novel techniques for acoustic-source location for both actively triggered, and passively detected signals using pervasive, distributed networks of devices, and investigates the combination of existing resources available in personal electronics to build a digital sensing 'commons'. By connecting personal resources with those of the people nearby, tasks can be achieved, through distributed placement and statistical improvement, that a single device could not do alone. The utility and benefits of spatio-temporal acoustic sensing are presented, in the context of ubiquitous computing and machine listening history. An active audio self-localisation algorithm is described which is effective in distributed sensor networks even if only coarse temporal synchronisation can be established. Pseudo-noise 'chirps' are emitted and recorded at each of the nodes. Pair-wise distances are calculated by comparing the difference in the audio delays between the peaks measured in each recording. By removing dependence on fine grained temporal synchronisation it is hoped that this technique can be used concurrently across a wide range of devices to better leverage the existing audio sensing resources that surround us.

(cont.) A passive acoustic source location estimation method is then derived which is suited to the microphone resources of network-connected heterogeneous devices containing asynchronous processors and uncalibrated sensors. Under these constraints position coordinates must be simultaneously determined for pairs of sounds and recorded at each microphone to form a chain of acoustic events. It is shown that an iterative, numerical least-squares estimator can be used. Initial position estimates of the source pair can be first found from the previous estimate in the chain and a closed-form least squares approach, improving the convergence rate of the second step. Implementations of these methods using the Smart Architectural Surfaces development platform are described and assessed. The viability of the active ranging technique is further demonstrated in a mixed-device ad-hoc sensor network case using existing off-the-shelf technology. Finally, drawing on human-centric onset detection as a means of discovering suitable sound features, to be passed between nodes for comparison, the extension of the source location algorithm beyond the use of pseudo-noise test sounds to enable the location of extraneous noises and acoustic streams is discussed for further study.