Presentation of the project

Fast Source Location Method UsingSearching Space Pre-estimation

EE627 Term PaperBatch 13

7 April 2015

IntroductionSource localization is a trending topic. More intelligent electronic devices are needed for office, home, and personal applications. For example • Video conferences- the camera is required to automatically

steer to the speaker of interest.• The hands-free operation of in-car electronics is necessary

during high-speed driving to avoid accidents.• Multiple signal classification (MUSIC) and Time difference

of arrival (TDOA) based location methods are two kinds of source localization techniques.

MUSIC- Multiple Signal Classification• It performs eigenvalue decomposition (EVD) of a covariance

matrix to obtain signal and noise subspace eigenvectors, then searches for location parameters that orthogonalize the steering vector and the noise subspace.

• Requires a search of whole region of interest (ROI), which is supposed to be computationally expensive.

• The main advantage of MUSIC lies in the fine spatial resolution but as short searching step must be assigned to obtain a good performance, which will be quite complex and require the calculation of a large number of steering vectors of the searching areas.

• In a reflective environment, MUSIC sometimes has a tendency to show a wrong estimate that is not true source direction, but the reflection’s.

TDOA- Time Difference Of Arrival• TDOAs are computed from different microphone pairs, and

then the individual TDOAs are combined to estimate the source location.

• Many TDOA estimation techniques have been proposed among which generalized eigenvalue decomposition (GEVD) and acoustic transfer functions ratio (ATF-s ratio) method exhibit the most robust estimate results under reverberant environment.

• ATF-s ratio algorithm derives TDOAs in frequency domain, thus shows comparatively low computational costs.

• A major defect of this kind of locators is that its precision counts on sampling rate of microphone signals.

TDOA+MUSIC: a better approach• First produces reliable areas by robust and efficient TDOA

estimation method, and then performs MUSIC algorithm within these areas.

• Such method not only owns the super resolution of MUSIC algorithm, but also avoids blindly testing all the directions belonging to the region of interest.

• Besides, such approach enhances the robustness of location estimation in reverberant conditions.

TDOA+MUSIC: a better approach

Block diagram illustrating TDOA+MUSIC

Data Model and Location Algorithm• Consider an array of M microphones. For a single source

located within the room, the received signal at the i-th microphone is given as

• is the source signal, is the impulse response between the source and i-th microphone and, is the additive noise at i-th microphone.

• Let be the transfer function of , then we define as the acoustic transfer function (ATF) ratio.

Data Model and Location Algorithm• Since signals observed at each microphone usually involve

the direct signal, as well as their reflected replicas, the transfer function will be

• where and denote the amplitudes and the delays of direct path and reflections in .

• Then the ATF ratio can be written as

Data Model and Location Algorithm• At low reverberation is close to 1. This means that the

TDOA is contained in the phase part of the ATF ratio.• Hence, the corresponding impulse response will have its

peak value at the sample delay between the i-th and 1-st microphone.

• The ATF-s ratio can be estimated from cross-PSD divided by auto-PSD,

where is the speech PSD.

Data Model and Location Algorithm• With the far field assumption, source bearing angle could

be coarsely computed according to achieved TDOA by

• is TDOA of signals between i-th and j-th microphone, is the angle of source direction with respect to array line.

• Having estimated the TDOA , the searching space for MUSIC* can be obtained by evaluating over ], where is the sampling time.

* MUSIC algorithm is discussed in detail in Schmidt, Ralph O. "Multiple emitter location and signal parameter estimation." Antennas and Propagation, IEEE Transactions on 34.3 (1986): 276-280.

Experiment and Simulation• A single sinusoidal source of frequency 1 KHz is incident on

a uniform linear array (ULA).• The ULA consists of 10 microphone elements having 0.5

spacing, where is the source wavelength.• The noise is assumed to be additive white Gaussian.• The SNR is greater than 10dB.• Reverberation effects are ignored in this experiment. • The source location is varied from 45 to 135 degree.

Experimental ResultsImpulse response of the ATF-ratio for DOA = 75 deg

Experimental Results

(a) MUSIC (b) MUSIC+TDOA

DOA=120

DOA=75DOA=75

DOA=120

Performance EvaluationRoot mean squared error of the estimated DOA, and runtime for 500 iterations

Conclusions• In this project, a fast MUSIC algorithm is presented by

making use of ATF ratio time delay estimation to determine the searching space.

• The TDOA+MUSIC algorithm is faster than the MUSIC algorithm (more than 4 times). However, the MUSIC algorithm is more accurate especially for DOAs closer to the array line.

• We can further investigate the performance of both algorithms under reverberant condition and source tracking scenario.