M.Sc Thesis


M.Sc StudentZeitouni Eran
SubjectSingle-Sensor Localization of Moving Sources using Diffusion
Kernels
DepartmentDepartment of Electrical and Computer Engineering
Supervisor PROF. Israel Cohen
Full Thesis textFull thesis text - English Version


Abstract

Source localization is a common problem in various fields and has applications in both military and civil sectors. Localization of acoustic sources generally requires a few microphones, but it is also possible to use a single microphone and data that was prerecorded in the same environment. Unfortunately, existing single microphone localization methods are restricted to acoustic sources that have a fixed location.

In this research, we introduce a supervised method for estimating both the location and velocity of a moving acoustic source, using a single microphone, based on a manifold learning approach. Each source, one at a time, transmits a white Gaussian noise signal during its movement, which is received by the sensor. The received signal is divided into time frames. The trajectory, formed by the movement of the source during the frame, is approximated by a linear movement segment. Each frame is inspected individually for estimation of the location and velocity. The signal, received during the frame, is processed to extract a feature vector. In such a manner, we collect a training set of observations, generated from different sources. Using diffusion maps with a Euclidean distance-based diffusion kernel, we learn the nonlinear structure of the manifold of the data. The data is organized on the manifold according to the location and velocity values of the sources. The unknown location and velocity of a source can be recovered according to its observation's nearest training neighbors on the manifold. The recovery of the location and velocity is determined by the midway point of the segment. Based on that point, we minimize the approximation error of the true trajectory by the linear segment. The performance of the proposed method is examined by various simulations investigating the sensitivity to different hyperparameters (e.g., frames length), variables (e.g., speed and direction), and conditions (e.g., reverberation time).

Research findings indicate that localization of very slow sources results in good accuracy of the estimated location, at the expense of relatively low accuracy of the estimated direction. Localization of faster sources leads to an improvement of the accuracy of the estimated direction due to a more meaningful movement of the sources, at the expense of the accuracy of the estimated location. The accuracy of the estimated direction starts deteriorating for fast sources, as the variations in their movement are too fast to be distinguished- resulting in a sparse manifold. The approximation by a linear segment is inaccurate for long frames, unless the source has constant velocity during the frame. Short frames are not good either, regardless the trajectory, as they are unable to capture the movement properly. The algorithm performs well in reverberant and noisy environments, yet is sensitive to environmental conditions changes. The results validate the necessity of reflections for yielding an accurate estimation. While the algorithm is designated for recovering the location and velocity of slow sources that change direction and speed gradually, it is even capable of successfully estimating the location and the average velocity of slow sources that change their velocity rapidly and randomly.