Visual / acoustic detection and localisation in embedded systems

Abstract

The continuous miniaturisation of sensing and processing technologies is increasingly offering a variety of embedded platforms, enabling the accomplishment of a broad range of tasks using such systems. Motivated by these advances, this thesis investigates embedded detection and localisation solutions using vision and acoustic sensors. Focus is particularly placed on surveillance applications using sensor networks. Existing vision-based detection solutions for embedded systems suffer from the sensitivity to environmental conditions. In the literature, there seems to be no algorithm able to simultaneously tackle all the challenges inherent to real-world videos. Regarding the acoustic modality, many research works have investigated acoustic source localisation solutions in distributed sensor networks. Nevertheless, it is still a challenging task to develop an ecient algorithm that deals with the experimental issues, to approach the performance required by these systems and to perform the data processing in a distributed and robust manner. The movement of scene objects is generally accompanied with sound emissions with features that vary from an environment to another. Therefore, considering the combination of the visual and acoustic modalities would offer a significant opportunity for improving the detection and/or localisation using the described platforms. In the light of the described framework, we investigate in the first part of the thesis the use of a cost-effective visual based method that can deal robustly with the issue of motion detection in static, dynamic and moving background conditions. For motion detection in static and dynamic backgrounds, we present the development and the performance analysis of a spatio- temporal form of the Gaussian mixture model. On the other hand, the problem of motion detection in moving backgrounds is addressed by accounting for registration errors in the captured images. By adopting a robust optimisation technique that takes into account the uncertainty about the visual measurements, we show that high detection accuracy can be achieved. In the second part of this thesis, we investigate solutions to the problem of acoustic source localisation using a trust region based optimisation technique. The proposed method shows an overall higher accuracy and convergence improvement compared to a linear-search based method. More importantly, we show that through characterising the errors in measurements, which is a common problem for such platforms, higher accuracy in the localisation can be attained. The last part of this work studies the different possibilities of combining visual and acoustic information in a distributed sensors network. In this context, we first propose to include the acoustic information in the visual model. The obtained new augmented model provides promising improvements in the detection and localisation processes. The second investigated solution consists in the fusion of the measurements coming from the different sensors. An evaluation of the accuracy of localisation and tracking using a centralised/decentralised architecture is conducted in various scenarios and experimental conditions. Results have shown the capability of this fusion approach to yield higher accuracy in the localisation and tracking of an active acoustic source than by using a single type of data.