track_audio.m

clear;
close all;
clc;

fs = 24414;
chunks = 2048;
dt = chunks/fs;

[Bf, Af] = butter(3,8000/fs); % filter vectors


%% Setup objects
% Initialize localization models using braodband and subband settings
dObj   = dataObject([],fs,10,2);

% Settings for subband approach
par_sub = genParStruct('cc_bBroadband',0,'cc_wSizeSec',winSec,...
    'cc_hSizeSec',winSec/2,'cc_maxDelaySec',1.25e-3,...
    'fb_lowFreqHz',fLowHz,'fb_highFreqHz',fHighHz,...
    'fb_nERBs',1,'ihc_method','none',...
    'loc_NSources',nSpeakers(hh));

% Initialize localization models using braodband and subband settings
mObj  = manager(dObj,'localization',par_sub);

%% Model parameters
sigma_w = 1;

Q = [2/4*dt^4, 1/2*dt^3; 1/2*dt^3, dt^2]*sigma_w;   % Process noise covariance
R = 1;                                              % Measurement noise covariance
x = [0; 0];                                         % Initial state
P = [10, 0; 0, 10];                                 % Initial state covariance

A = [1, dt; 0, 1];                                  % System matrix
c = [1; 0];                                         % Output vector

% Check definiteness of covariance matrices
if ~all(eig(Q) > 0) || ~all(eig(R) > 0) || ~all(eig(A) > 0)
    error('All covariance matrices have to be positive definite.');
end

%% Initialization


% Add necessary paths
addpath('./tools');
addpath('./ekfukf-toolbox');



figure(1)

N = 1; % The number of steps to run this stuff.

% Initialize posterior mean and covariance
posteriorMean = zeros(size(A, 1), N);
posteriorCovariance = zeros(size(A, 1), size(A, 1), N);


% =======================================================
% Main loop - Perform localization and tracking
% =======================================================
tic();
t_old = toc();
for l = 1:N
    audio = get_audio();
    t_new = toc();
    dti = t_new - t_old();
    % Request processing
    mObj.processSignal(audio);
    azimEst = dObj.localization{1}.Data(end,1); % There might be an issue with several sources here!
    % Perform Kalman filter prediction and update, TODO: consider changing this
    % crappy filter for a PF
    
    
    Qi = [1/4*dti^4+1e-6, 1/2*dti^3; 1/2*dti^3, dti^2]*sigma_w;   % Process noise covariance
    Ai = [1, dti; 0, 1];
    [x, P] = kf_predict(x, P, Ai, Qi);
    [x, P] = kf_update(x, P, azimEst, c', R);
    
    posteriorMean(:, l) = x;
    posteriorCovariance(:, :, l) = P;
    %pause(max0))
    t_old = t_new;
end

% Plot measurements
subplot(2, nFiles / 2, k);
timeAxis = linspace(0, nSamples / fsHz, nFrames);
plot(timeAxis, measuredLocations, 'x', 'LineWidth', 2);
axis([0, nSamples / fsHz, -90, 90]);
xlabel('Time / s');
ylabel('Azimuth / deg');
grid on; hold on;
plot(timeAxis, posteriorMean(1, :), 'g', 'LineWidth', 2);

% Plot ground truth
plot(timeAxis, gtTrajectory, 'r--', 'LineWidth', 2);
legend('Measurements', 'Estimated trajectory', 'Ground truth');

% Compute RMSE
rmse = sqrt(sum((posteriorMean(1, :) - gtTrajectory).^2) ./ nFrames);

if ~strcmpi(noiseType, 'none')
    title([upper(soundType), ', ', upper(noiseType), ' NOISE AT ', ...
        num2str(snr), ' dB SNR, ', 'RMSE: ', num2str(rmse), '°']);
else
    title([upper(soundType), ', NO NOISE, ', 'RMSE: ', ...
        num2str(rmse), '°']);
end