toni/Dirigent
1
0
Fork 0
Code Issues 17 Pull Requests Releases 2 Wiki Activity
Files
85ea37c14bf7bf37ada1c593569bb8a5f40356f4
Dirigent/matlab/AutoCorrMethodNew_Watch.m
toni 85ea37c14b ref #27 -new gradle and sdk version 
- added faster bpm estimation by using multiple windows now
- implemented a simple z score based outlier detection
- some ugly heuristics to prefend to many -1 on the screen
NOTE: this is a very unstable version of the code, very protoype
2018-04-27 16:03:58 +02:00

312 lines
14 KiB
Matlab
Raw Blame History

%We are using a threshold-based version for bpm estimation
%Only the z axis of the acc is used
%NOTE: depending on the measurement device we have a highly different sample rate. the smartwatches are not capable of providing a constant sample rate. the xsens on the other hand is able to do this.
%load file provided by the sensor readout app
% SMARTWATCH LG WEAR ------> 100 hz - 1000hz
%measurements = dlmread('../../measurements/lgWear/PR_recording_80bpm_4-4_177596720.csv', ';'); %
%measurements = dlmread('../../measurements/lgWear/recording_48bpm_4-4_176527527.csv', ';');
%measurements = dlmread('../../measurements/lgWear/recording_48bpm_4-4_176606785.csv', ';');
%measurements = dlmread('../../measurements/lgWear/recording_48bpm_4-4_176696356.csv', ';');
%measurements = dlmread('../../measurements/lgWear/recording_48bpm_4-4_176820066.csv', ';'); %
%measurements = dlmread('../../measurements/lgWear/recording_48bpm_4-4_176931941.csv', ';'); %double
%measurements = dlmread('../../measurements/lgWear/recording_72bpm_4-4_176381633.csv', ';');
%measurements = dlmread('../../measurements/lgWear/recording_72bpm_4-4_176453327.csv', ';'); %
%measurements = dlmread('../../measurements/lgWear/recording_100bpm_4-4_176073767.csv', ';'); %*
%measurements = dlmread('../../measurements/lgWear/recording_100bpm_4-4_176165357.csv', ';');
%measurements = dlmread('../../measurements/lgWear/recording_100bpm_4-4_176230146.csv', ';');
%measurements = dlmread('../../measurements/lgWear/recording_100bpm_4-4_176284687.csv', ';'); %
%measurements = dlmread('../../measurements/lgWear/recording_100bpm_4-4_177368860.csv', ';'); %(besonders)
%measurements = dlmread('../../measurements/lgWear/recording_180bpm_4-4_177011641.csv', ';'); %
%measurements = dlmread('../../measurements/lgWear/recording_180bpm_4-4_177064915.csv', ';'); %
% SMARTWATCH G WATCH WEAR R ----> 100hz - 250hz
%measurements = dlmread('../measurements/wearR/PR_recording_80bpm_4-4_177596720.csv', ';'); %*
%measurements = dlmread('../measurements/wearR/recording_48bpm_4-4_176527527.csv', ';');
%measurements = dlmread('../measurements/wearR/recording_48bpm_4-4_176606785.csv', ';');
%measurements = dlmread('../../measurements/wearR/recording_48bpm_4-4_176696356.csv', ';'); %*
%measurements = dlmread('../measurements/wearR/recording_48bpm_4-4_176820066.csv', ';');
%measurements = dlmread('../measurements/wearR/recording_48bpm_4-4_176931941.csv', ';'); %double
%measurements = dlmread('../measurements/wearR/recording_72bpm_4-4_176381633.csv', ';');
%measurements = dlmread('../measurements/wearR/recording_72bpm_4-4_176453327.csv', ';');
%measurements = dlmread('../measurements/wearR/recording_100bpm_4-4_176073767.csv', ';'); *
%measurements = dlmread('../measurements/wearR/recording_100bpm_4-4_176165357.csv', ';');
%measurements = dlmread('../measurements/wearR/recording_100bpm_4-4_176230146.csv', ';'); %* 72?
%measurements = dlmread('../measurements/wearR/recording_100bpm_4-4_176284687.csv', ';'); %*48?
%measurements = dlmread('../measurements/wearR/recording_100bpm_4-4_177368860.csv', ';'); %(besonders)
%measurements = dlmread('../measurements/wearR/recording_180bpm_4-4_177011641.csv', ';'); *
%measurements = dlmread('../measurements/wearR/recording_180bpm_4-4_177064915.csv', ';'); *
files = dir(fullfile('../../measurements/lgWear/', '*.csv'));
%files = dir(fullfile('../../measurements/wearR/', '*.csv'));
%files = dir(fullfile('../../measurements/peter_failed/', '*.csv'));
for file = files'
filename = [file.folder '/' file.name];
measurements = dlmread(filename, ';');
%draw the raw acc data
m_idx = [];
m_idx = (measurements(:,2)==2); %Android App: 10, Normal Data: 2
m = measurements(m_idx, :);
%Interpolate to generate a constant sample rate to 250hz (4ms per sample)
sample_rate_ms = 4;%ms
[~, m_unique_idx] = unique(m(:,1)); %matlab requirs unique timestamps for interp
m = m(m_unique_idx, :);
t = m(:,1); %timestamps
t_interp = t(1):sample_rate_ms:t(length(t));
m_interp = interp1(t,m(:,3:5),t_interp);
%put all together again
m = [t_interp', t_interp', m_interp];
figure(1);
plot(m(:,1),m(:,3)) %x
legend("x", "location", "eastoutside");
figure(2);
plot(m(:,1),m(:,4)) %yt
legend("y", "location", "eastoutside");
figure(3);
plot(m(:,1),m(:,5)) %z
legend("z", "location", "eastoutside");
%magnitude
magnitude = sqrt(sum(m(:,3:5).^2,2));
figure(5);
plot(m(:,1), magnitude);
legend("magnitude", "location", "eastoutside");
waitforbuttonpress();
%save timestamps
timestamps = m(:,1);
data = m(:,3); %only z
%TODO: Different window sizes for periods under 16.3 s
window_size = 2048; %about 2 seconds using 2000hz, 16.3 s using 250hz
overlap = 256;
bpm_per_window_ms = [];
bpm_per_window = [];
for i = window_size+1:length(data)
%wait until window is filled with new data
if(mod(i,overlap) == 0)
%measure periodicity of window and use axis with best periodicity
[corr_x, lag_x] = xcov(m(i-window_size:i,3), (window_size/4), "coeff");
[corr_y, lag_y] = xcov(m(i-window_size:i,4), (window_size/4), "coeff");
[corr_z, lag_z] = xcov(m(i-window_size:i,5), (window_size/4), "coeff");
[corr_mag, lag_mag] = xcov(magnitude(i-window_size:i), (window_size/4), "coeff");
%autocorrelation of the autocorrelation?!
%[corr_corr_x, lag_lag_x] = xcov(corr_x, length(corr_x), "coeff");
%[corr_corr_y, lag_lag_y] = xcov(corr_y, length(corr_x), "coeff");
%[corr_corr_z, lag_lag_z] = xcov(corr_z, length(corr_x), "coeff");
corr_x_pos = corr_x;
corr_y_pos = corr_y;
corr_z_pos = corr_z;
corr_mag_pos = corr_mag;
corr_x_pos(corr_x_pos<0)=0;
corr_y_pos(corr_y_pos<0)=0;
corr_z_pos(corr_z_pos<0)=0;
corr_mag_pos(corr_mag_pos<0)=0;
[peak_x, idx_x_raw] = findpeaks(corr_x_pos, 'MinPeakHeight', 0.1,'MinPeakDistance', 50, 'MinPeakProminence', 0.1);
[peak_y, idx_y_raw] = findpeaks(corr_y_pos, 'MinPeakHeight', 0.1,'MinPeakDistance', 50, 'MinPeakProminence', 0.1);
[peak_z, idx_z_raw] = findpeaks(corr_z_pos, 'MinPeakHeight', 0.1,'MinPeakDistance', 50, 'MinPeakProminence', 0.1);
[peak_mag, idx_mag_raw] = findpeaks(corr_mag_pos, 'MinPeakHeight', 0.1,'MinPeakDistance', 50, 'MinPeakProminence', 0.1);
idx_x_raw = sort(idx_x_raw);
idx_y_raw = sort(idx_y_raw);
idx_z_raw = sort(idx_z_raw);
idx_mag_raw = sort(idx_mag_raw);
idx_x = findFalseDetectedPeaks(idx_x_raw, lag_x, corr_x);
idx_y = findFalseDetectedPeaks(idx_y_raw, lag_y, corr_y);
idx_z = findFalseDetectedPeaks(idx_z_raw, lag_z, corr_z);
idx_mag = findFalseDetectedPeaks(idx_mag_raw, lag_mag, corr_mag);
Xwindow = m(i-window_size:i,3);
Xwindow_mean_ts_diff = mean(diff(lag_x(idx_x) * sample_rate_ms)); %2.5 ms is the time between two samples at 400hz
Xwindow_mean_bpm = (60000 / (Xwindow_mean_ts_diff));
figure(11);
plot(lag_x, corr_x, lag_x(idx_x), corr_x(idx_x), 'r*', lag_x(idx_x_raw), corr_x(idx_x_raw), 'g*') %z
hold ("on")
m_label_ms = strcat(" mean ms: ", num2str(Xwindow_mean_ts_diff));
m_label_bpm = strcat(" mean bpm: ", num2str(Xwindow_mean_bpm));
title(strcat(" ", m_label_ms, " ", m_label_bpm));
hold ("off");
Ywindow = m(i-window_size:i,4);
Ywindow_mean_ts_diff = mean(diff(lag_y(idx_y) * sample_rate_ms));
Ywindow_mean_bpm = (60000 / (Ywindow_mean_ts_diff));
figure(12);
plot(lag_y, corr_y, lag_y(idx_y), corr_y(idx_y), 'r*', lag_y(idx_y_raw), corr_y(idx_y_raw), 'g*') %z
hold ("on")
m_label_ms = strcat(" mean ms: ", num2str(Ywindow_mean_ts_diff));
m_label_bpm = strcat(" mean bpm: ", num2str(Ywindow_mean_bpm));
title(strcat(" ", m_label_ms, " ", m_label_bpm));
hold ("off");
Zwindow = m(i-window_size:i,5);
Zwindow_mean_ts_diff = mean(diff(lag_z(idx_z)* sample_rate_ms));
Zwindow_mean_bpm = (60000 / (Zwindow_mean_ts_diff));
figure(13);
plot(lag_z, corr_z, lag_z(idx_z), corr_z(idx_z), 'r*', lag_z(idx_z_raw), corr_z(idx_z_raw), 'g*') %z
hold ("on")
m_label_ms = strcat(" mean ms: ", num2str(Zwindow_mean_ts_diff));
m_label_bpm = strcat(" mean bpm: ", num2str(Zwindow_mean_bpm));
title(strcat(" ", m_label_ms, " ", m_label_bpm));
hold ("off");
%magnitude
Mwindow = magnitude(i-window_size:i);
Mwindow_mean_ts_diff = mean(diff(lag_mag(idx_mag)* sample_rate_ms));
Mwindow_mean_bpm = (60000 / (Mwindow_mean_ts_diff));
figure(14);
plot(lag_mag, corr_mag, lag_mag(idx_mag), corr_mag(idx_mag), 'r*', lag_mag(idx_mag_raw), corr_mag(idx_mag_raw), 'g*') %z
hold ("on")
m_label_ms = strcat(" mean ms: ", num2str(Mwindow_mean_ts_diff));
m_label_bpm = strcat(" mean bpm: ", num2str(Mwindow_mean_bpm));
title(strcat(" ", m_label_ms, " ", m_label_bpm));
hold ("off");
%breakpoints dummy for testing
if(length(idx_x) > length(idx_x_raw))
a = 0; %breakpointdummy
end
if(length(idx_y) > length(idx_y_raw))
a = 0; %breakpointdummy
end
if(length(idx_z) > length(idx_z_raw))
a = 0; %breakpointdummy
end
%Find the most proper axis. We use 3 quantities: mean of corr.
%value, sum of corr val. and number of peaks. Simple normalization
%to get the axis that fullfills the quantities the most.
idx_noZero_x = idx_x(lag_x(idx_x) ~= 0);
idx_noZero_y = idx_y(lag_x(idx_y) ~= 0);
idx_noZero_z = idx_z(lag_x(idx_z) ~= 0);
corr_mean_x = geomean(corr_x(idx_noZero_x(corr_x(idx_noZero_x)>0)));
corr_mean_y = geomean(corr_y(idx_noZero_y(corr_y(idx_noZero_y)>0)));
corr_mean_z = geomean(corr_z(idx_noZero_z(corr_z(idx_noZero_z)>0)));
corr_rms_x = rms(corr_x(idx_x(lag_x(idx_x) ~= 0)));
corr_rms_y = rms(corr_y(idx_y(lag_y(idx_y) ~= 0)));
corr_rms_z = rms(corr_z(idx_z(lag_z(idx_z) ~= 0)));
num_peaks_x = 1;%length(idx_x);
num_peaks_y = 1;%length(idx_y);
num_peaks_z = 1;%length(idx_z);
num_intersection_x = getNumberOfIntersections(corr_x, lag_x, 0.2);
num_intersection_y = getNumberOfIntersections(corr_y, lag_y, 0.2);
num_intersection_z = getNumberOfIntersections(corr_z, lag_z, 0.2);
quantity_matrix = [corr_mean_x corr_mean_y corr_mean_z;
corr_rms_x corr_rms_y corr_rms_z;
num_intersection_x num_intersection_y num_intersection_z];
quantity_matrix_percent(1,:) = quantity_matrix(1,:) ./ sum(quantity_matrix(1,:));
quantity_matrix_percent(2,:) = quantity_matrix(2,:) ./ sum(quantity_matrix(2,:));
quantity_matrix_percent(3,:) = quantity_matrix(3,:) ./ sum(quantity_matrix(3,:));
quantity_factors = sum(quantity_matrix_percent) / 3;
%TODO: Wenn ein quantity wert NaN ist, sind alle NaN...
quantity_x = quantity_factors(1);
quantity_y = quantity_factors(2);
quantity_z = quantity_factors(3);
%choose axis with sum(corr) nearest to 0
%{
corr_sum_xyz = [sum(corr_x) sum(corr_y) sum(corr_z)];
[~,idx_nearest_zero] = min(abs(corr_sum_xyz));
if(idx_nearest_zero == 1)
window_mean_ts_diff = Xwindow_mean_ts_diff;
window_mean_bpm = Xwindow_mean_bpm;
elseif(idx_nearest_zero == 2)
window_mean_ts_diff = Ywindow_mean_ts_diff;
window_mean_bpm = Ywindow_mean_bpm;
else
window_mean_ts_diff = Zwindow_mean_ts_diff;
window_mean_bpm = Zwindow_mean_bpm;
end
%}
%quantity_x = num_intersection_x;
%quantity_y = num_intersection_y;
%quantity_z = num_intersection_z;
if(quantity_x > quantity_y && quantity_x > quantity_z)
window_mean_ts_diff = Xwindow_mean_ts_diff;
window_mean_bpm = Xwindow_mean_bpm;
elseif(quantity_y > quantity_z)
window_mean_ts_diff = Ywindow_mean_ts_diff;
window_mean_bpm = Ywindow_mean_bpm;
else
window_mean_ts_diff = Zwindow_mean_ts_diff;
window_mean_bpm = Zwindow_mean_bpm;
end
if(isnan(window_mean_ts_diff) || isnan(window_mean_bpm))
%do nothing
else
bpm_per_window_ms = [bpm_per_window_ms, window_mean_ts_diff];
bpm_per_window = [bpm_per_window, window_mean_bpm];
end
%TODO: if correlation value is lower then a treshhold, we are not conducting TODO: change to a real classification instead of a treshhold.
end
end
%TODO: smooth the results using a moving avg or 1d kalman filter.(transition for kalman could be adding the last measured value)
%remove the first 40% of the results, due to starting delays while recording.
number_to_remove = round(abs(0.1 * length(bpm_per_window_ms)));
num_all = length(bpm_per_window_ms);
bpm_per_window_ms = bpm_per_window_ms(number_to_remove:num_all);
bpm_per_window = bpm_per_window(number_to_remove:num_all);
mean_final_ms = mean(bpm_per_window_ms);
std_final_ms = std(bpm_per_window_ms);
mean_final_bpm = mean(bpm_per_window);
std_final_bpm = std(bpm_per_window);
fprintf('%s: mean = %f bpm (%f ms) stddev = %f bpm (%f ms)\n', strrep(regexprep(filename,'^.*recording_',''),'.txt',''), mean_final_bpm, mean_final_ms, std_final_bpm, std_final_ms);
end
Reference in New Issue View Git Blame Copy Permalink