%using autocorrelation to estimate the current bmp within some fixed window %load sensor files %files = dir(fullfile('../../measurements/2017.06/mSensor/', '*.csv')); %files = natsortfiles(dir(fullfile('../../measurements/2017.06/lgWear/', '*.csv'))); %files = dir(fullfile('../../measurements/wearR/', '*.csv')); %files = dir(fullfile('../../measurements/peter_failed/', '*.csv')); %files = dir(fullfile('../../measurements/2018.06/manfred/LGWatchR/', '*.csv')); %files = dir(fullfile('../../measurements/2018.06/peter/Huawai/', '*.csv')); %files = dir(fullfile('../../measurements/2018.06/peter/mSensor/', '*.csv')); %files = dir(fullfile('../../measurements/2018.06/frank/mSensor/', '*.csv')); files = dir(fullfile('../../measurements/2018.06/leon/mSensor/', '*.csv')); %files_sorted = natsortfiles({files.name}); for file = files' filename = [file.folder '/' file.name]; measurements = dlmread(filename, ';', 3, 0); %load ground truth file fid = fopen(filename); fgetl(fid); Str = fgetl(fid); Key = 'Metronom: '; Index = strfind(Str, Key); gtDataRaw = sscanf(Str(Index(1) + length(Key):end), '%g', 1); gtData = []; gtFile = []; if(isempty(gtDataRaw)) gtFile = extractAfter(Str, Key); gtFile = strcat('../../measurements/2018.06/gt_toni/', gtFile); f = fopen(gtFile); gtDataRaw = textscan(f, '%f %s', 'Delimiter', ' '); fclose(f); [~,~,~,hours,minutes,seconds] = datevec(gtDataRaw{2}, 'HH:MM:SS.FFF'); gtData(:,1) = 1000*(60*minutes + seconds); %we do not use hours! gtData(:,2) = gtDataRaw{1}; else gtData = gtDataRaw; end %draw the raw acc data m_idx = []; m_idx = (measurements(:,2)==3); %Android App: 10, Sensor: 3, 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 = 1024; %about 2 seconds using 2000hz, 16.3 s using 250hz overlap = 256; bpm_per_window_ms = []; bpm_per_window = []; bpm_3D = []; ms_3D = []; gtIdx = 1; gtError_3D = []; gtError_1D = []; for i = window_size+1:1:length(data) %wait until window is filled with new data if(mod(i,overlap) == 0) %set cur ground truth if(length(gtData) > 1) curTimestamp = timestamps(i); while(curTimestamp > gtData(gtIdx,1) && gtIdx < length(gtData)) curGtBpm = gtData(gtIdx,2); gtIdx = gtIdx + 1; end else curGtBpm = gtData; end %measure periodicity of window and use axis with best periodicity [corr_x, lag_x] = xcov(m(i-window_size:i,3), (window_size/2), "coeff"); [corr_y, lag_y] = xcov(m(i-window_size:i,4), (window_size/2), "coeff"); [corr_z, lag_z] = xcov(m(i-window_size:i,5), (window_size/2), "coeff"); %magnitude [corr_mag, lag_mag] = xcov(magnitude(i-window_size:i), (window_size/2), "coeff"); %TODO: stichwort spatial autocorrelation %figure(77); %scatter3(timestamps(i-window_size:i), m(i-window_size:i,4), m(i-window_size:i,5)); %distanz zwischen den vektoren nehmen und in eine normale autocorrelation zu packen %aufpassen wegen der norm, dass die richtung quasi nicht verloren geht. %https://en.wikipedia.org/wiki/Lp_space [corr_3D, lag_3D] = distCorr(m(i-window_size:i, 3:5), (round(window_size * 0.8))); 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); [peak_3D, idx_3D_raw] = findpeaks(corr_3D, '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_3D_raw = sort(idx_3D_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); idx_3D = findFalseDetectedPeaks(idx_3D_raw, lag_3D', corr_3D); Dwindow = m(i-window_size:i,3); Dwindow_mean_ts_diff = mean(diff(lag_3D(idx_3D) * sample_rate_ms)); %2.5 ms is the time between two samples at 400hz Dwindow_mean_bpm = (60000 / (Dwindow_mean_ts_diff)); figure(10); plot(lag_3D, corr_3D, lag_3D(idx_3D), corr_3D(idx_3D), 'r*', lag_3D(idx_3D_raw), corr_3D(idx_3D_raw), 'g*') hold ("on") m_label_ms = strcat(" mean ms: ", num2str(Dwindow_mean_ts_diff)); m_label_bpm = strcat(" mean bpm: ", num2str(Dwindow_mean_bpm)); title(strcat(" ", m_label_ms, " ", m_label_bpm)); hold ("off"); 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 gtError_1D = [gtError_1D, abs(window_mean_bpm - curGtBpm)]; bpm_per_window_ms = [bpm_per_window_ms, window_mean_ts_diff]; bpm_per_window = [bpm_per_window, window_mean_bpm]; end %3D mean if(isnan(Dwindow_mean_bpm)) %nothing else gtError_3D = [gtError_3D, abs(Dwindow_mean_bpm - curGtBpm)]; bpm_3D = [bpm_3D, Dwindow_mean_bpm]; ms_3D = [ms_3D, Dwindow_mean_ts_diff]; 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); mean_final_error_1D = mean(gtError_1D); std_final_error_1D = std(gtError_1D); mean_final_ms_3D = mean(ms_3D); std_final_ms_3D = std(ms_3D); mean_final_bpm_3D = mean(bpm_3D); std_final_bpm_3D = std(bpm_3D); mean_final_error_3D = mean(gtError_3D); std_final_error_3D = std(gtError_3D); fprintf('%s: mean = %f bpm (%f bpm) stddev = %f bpm (%f bpm) --- 1D\n', strrep(regexprep(filename,'^.*recording_',''),'.txt',''), mean_final_error_1D, mean_final_bpm, std_final_error_1D, std_final_bpm); fprintf('%s: mean = %f bpm (%f bpm) stddev = %f bpm (%f bpm) --- 3D\n', strrep(regexprep(filename,'^.*recording_',''),'.txt',''), mean_final_error_3D, mean_final_bpm_3D, std_final_error_3D, std_final_bpm_3D); end % %1D fft - nicht so der brüller % z_fft = fft(m(i-window_size:i,5)); % L = length(z_fft); % Fs = 250; % P2 = abs(z_fft/L); % P1 = P2(1:L/2+1); % P1(2:end-1) = 2*P1(2:end-1); % f = Fs*(0:(L/2))/L; %nyquist frequence % % figure(66); % plot(f, P1); % % %3D fft % m_3D = m(i-window_size:i, 3); % m_3D(:,:,2) = m(i-window_size:i, 4); % m_3D(:,:,3) = m(i-window_size:i, 5); % % fft_3D = fftn(m_3D); % % %2D fft % fft_xy = fft2(m(i-window_size:i, 3:4)); % fft_yz = fft2(m(i-window_size:i, 3:4)); % % fft_test = fft(m(i-window_size:i, 3:5),[],2); % % figure(60); % imagesc(abs(fftshift(fft_test))); % % figure(61); % imagesc(abs(fftshift(fft_xy)));