added classification, results are bad...

toni/octave/functions.m

@@ -1,8 +1,10 @@
 display("functions")
+source("settings.m");
+
 
 function files = getDataFiles(clsName, trainsetPerClass)
-  
+
-  setDataDir = "/home/toni/Documents/handygames/HandyGames/daten/";
+  global setDataDir;
   
   dir = strcat(setDataDir, clsName, "/");
   lst = readdir(dir);
@@ -36,13 +38,16 @@ function samples = getSamplesForClass(clsName, trainsetPerClass, start, percent)
   display(strcat("training data for '", clsName, "': ", num2str(length(samples)), " samples"));
 end
 
-function data = getRawTrainData(trainsetPerClass)
+function data = getRawTrainData()
+  
+  #global trainsetPerClass;
+
   data = {};
-  data{1}.samples = getSamplesForClass("forwardbend", trainsetPerClass, 10, 0.9);
+  data{1}.samples = getSamplesForClass("forwardbend", 9, 10, 0.95);
-  data{2}.samples = getSamplesForClass("kneebend", trainsetPerClass, 10, 0.9);
+  data{2}.samples = getSamplesForClass("kneebend", 11, 10, 0.95);
-  data{3}.samples = getSamplesForClass("pushups", trainsetPerClass, 10, 0.9);
+  data{3}.samples = getSamplesForClass("pushups", 11, 10, 0.95);
-  data{4}.samples = getSamplesForClass("situps", trainsetPerClass, 10, 0.9);
+  data{4}.samples = getSamplesForClass("situps", 8, 10, 0.95);
-  data{5}.samples = getSamplesForClass("jumpingjack", trainsetPerClass, 10, 0.9);
+  data{5}.samples = getSamplesForClass("jumpingjack", 13, 10, 0.95);
 end
 
 function plotData(data,outPath)
@@ -75,17 +80,15 @@ function filteredData = filterData(data)
 end
 
 function win = window(rawVec, posMS)
+  
+  global setWindowSize;
   pos = posMS / 10;
-  #check if out of bounce, if yes- fill with zeros
+  
-  #the last windows are sometimes filled with a lot of zeros... this could cause problems
+  #only full windows are accepted. 
-  if length(rawVec) <= pos+100-1
+  if length(rawVec) <= pos+(setWindowSize/2)-1
-    #win = rawVec(pos-100:length(rawVec),:);
-    #fillOut = zeros((pos+100-1) - length(rawVec), 3);
-    #length(fillOut)
-    #win = [win; fillOut];
     win = [];
   else  
-    win = rawVec(pos-100:pos+100-1,:); #100*10 ms is half the window size. that is 2s for each window. 
+    win = rawVec(pos-(setWindowSize/2):pos+(setWindowSize/2)-1,:); #100*10 ms is half the window size. 
   endif
 end
 
@@ -100,14 +103,24 @@ function out = getMagnitude(data)
 end
 
 function windowedData = windowData(data)
-windowedData = {};
+
+  global setWindowSize;
+  global setWindowSliding;
+  windowedData = {};
 
   for k = 1:numel(data)
     for j = 1:numel(data{k}.samples)
     
-      winsAccel = {};
+      #init
-      winsGyro = {};
+      winsAccelX = {};
-      winsMagnet = {};
+      winsAccelY = {};
+      winsAccelZ = {};
+      winsGyroX = {};
+      winsGyroY = {};
+      winsGyroZ = {};
+      winsMagnetX = {};
+      winsMagnetY = {};
+      winsMagnetZ = {};
       
       winsAccelPCA = {};
       winsGyroPCA = {};
@@ -118,12 +131,15 @@ windowedData = {};
       winsMagnetMG = {};
       
       winEnd = length(data{k}.samples{j}.raw{1}) * 10; # *10ms 
+      winBuffer = ((setWindowSize/2)+1) * 10;
       
-      for i = 1010:200:winEnd-1010 #200ms steps for sliding/overlapping windows
+      for i = winBuffer:setWindowSliding:winEnd-winBuffer #steps for sliding/overlapping windows 
         
         #accel
         winAccel = window(data{k}.samples{j}.raw{1}, i);
-        winsAccel = [winsAccel winAccel];
+        winsAccelX = [winsAccelX winAccel(:,1)];
+        winsAccelY = [winsAccelY winAccel(:,2)];
+        winsAccelZ = [winsAccelZ winAccel(:,3)];
         
         [coeff1, score1] = princomp(winAccel);
         winsAccelPCA = [winsAccelPCA score1(:,1)];  #choose the first axis (eigvec with the biggest eigvalue)
@@ -133,7 +149,9 @@ windowedData = {};
         
         #gyro
         winGyro = window(data{k}.samples{j}.raw{2}, i);
-        winsGyro = [winsGyro winGyro];
+        winsGyroX = [winsGyroX winGyro(:,1)];
+        winsGyroY = [winsGyroY winGyro(:,2)];
+        winsGyroZ = [winsGyroZ winGyro(:,3)];
         
         [coeff2, score2] = princomp(winGyro);
         winsGyroPCA = [winsGyroPCA score2(:,1)];
@@ -143,7 +161,9 @@ windowedData = {};
         
         #magnet
         winMagnet = window(data{k}.samples{j}.raw{3}, i);
-        winsMagnet = [winsMagnet winMagnet];
+        winsMagnetX = [winsMagnetX winMagnet(:,1)];
+        winsMagnetY = [winsMagnetY winMagnet(:,2)];
+        winsMagnetZ = [winsMagnetZ winMagnet(:,3)];        
         
         [coeff3, score3] = princomp(winMagnet);
         winsMagnetPCA = [winsMagnetPCA score3(:,1)];
@@ -154,53 +174,68 @@ windowedData = {};
       end
       
       #write back data
-      windowedData{k}.samples{j}.raw{1}.wins = winsAccel;
+      windowedData{k}.samples{j}.raw{1}.wins = winsAccelX; #X
-      windowedData{k}.samples{j}.raw{2}.wins = winsGyro;
+      windowedData{k}.samples{j}.raw{2}.wins = winsAccelY; #Y
-      windowedData{k}.samples{j}.raw{3}.wins = winsMagnet; 
+      windowedData{k}.samples{j}.raw{3}.wins = winsAccelZ; #Z
-      windowedData{k}.samples{j}.raw{4}.wins = winsAccelPCA;
+      windowedData{k}.samples{j}.raw{4}.wins = winsGyroX;
-      windowedData{k}.samples{j}.raw{5}.wins = winsGyroPCA;
+      windowedData{k}.samples{j}.raw{5}.wins = winsGyroY;
-      windowedData{k}.samples{j}.raw{6}.wins = winsMagnetPCA;          
+      windowedData{k}.samples{j}.raw{6}.wins = winsGyroZ;
-      windowedData{k}.samples{j}.raw{7}.wins = winsAccelMG;
+      windowedData{k}.samples{j}.raw{7}.wins = winsMagnetX;
-      windowedData{k}.samples{j}.raw{8}.wins = winsGyroMG;
+      windowedData{k}.samples{j}.raw{8}.wins = winsMagnetY;
-      windowedData{k}.samples{j}.raw{9}.wins = winsMagnetMG;         
+      windowedData{k}.samples{j}.raw{9}.wins = winsMagnetZ; 
+      windowedData{k}.samples{j}.raw{10}.wins = winsAccelPCA;
+      windowedData{k}.samples{j}.raw{11}.wins = winsGyroPCA;
+      windowedData{k}.samples{j}.raw{12}.wins = winsMagnetPCA;          
+      windowedData{k}.samples{j}.raw{13}.wins = winsAccelMG;
+      windowedData{k}.samples{j}.raw{14}.wins = winsGyroMG;
+      windowedData{k}.samples{j}.raw{15}.wins = winsMagnetMG;         
     end
   end  
 end
 
 function features = featureCalculation(data)
 
-features = [];
+  global setAutocorrelationBinSize;
+  global setPSDBinSize;
+  features = [];
 
   for k = 1:numel(data)
     for j = 1:numel(data{k}.samples)
       for i = 1:numel(data{k}.samples{j}.raw)
         for m = 1:numel(data{k}.samples{j}.raw{i}.wins)
           currentWindow = data{k}.samples{j}.raw{i}.wins{m};
-          currentWindow = currentWindow(:,1);
           
           #autocorrelation on window. split into 5 evenly spaced bins (frequencies are evenly spaced, not number of values ;) ) and calculate mean of bin. 
           [autoCorr] = xcorr(currentWindow);
-          [binNum, binCenter] = hist(autoCorr, 5); #define 5 bins for the data.
+          [binNum, binCenter] = hist(autoCorr, setAutocorrelationBinSize); #define bins for the data.
           binSize = abs(binCenter(end-1) - binCenter(end));
-          binEdges = linspace(binCenter(1)-(binSize/2), binCenter(end)+(binSize/2), 6);
+          binEdges = linspace(binCenter(1)-(binSize/2), binCenter(end)+(binSize/2), setAutocorrelationBinSize+1);
           [binNumc, binIdx] = histc(autoCorr, binEdges);
-          binMeans = getBinMean(autoCorr, binIdx, 5);
+          binMeans = getBinMean(autoCorr, binIdx, setAutocorrelationBinSize);
           
           #calculate the root-mean-square (RMS) of the signal
           rms = sqrt(mean(currentWindow.^2));
           
           #power bands 0.5 to 25hz (useful if the windows are greater then 4s and window sizes to 256, 512..) 
           [powerBand, w] = periodogram(currentWindow); #fills up fft with zeros            
-          powerEdges = logspace(log10(0.5), log10(25), 10 + 2);  #logarithmic bin spaces for 10 bins
+          powerEdges = logspace(log10(0.5), log10(25), setPSDBinSize + 2);  #logarithmic bin spaces for 10 bins
           triFilter = getTriangularFunction(powerEdges, length(powerBand)*2 - 2);    
           for l = 1:numel(triFilter)
             filteredBand = triFilter{l} .* powerBand;
             psd(l) = sum(filteredBand); #sum freq (no log and no dct)
           end
           
+          #statistical features
+          windowMean = mean(currentWindow);
+          windowSTD = std(currentWindow); #standard deviation
+          windowVariance = var(currentWindow);
+          windowKurtosis = kurtosis(currentWindow); #(ger. Wölbung)
+          windowIQR = iqr(currentWindow); #interquartile range
+          
           #put everything together
           classLabel = k; #what class?
-          features = [features; classLabel, binMeans, rms, psd];
+          sampleLabel = j; #what sample?
+          features = [features; sampleLabel, classLabel, binMeans, rms, psd, windowMean, windowSTD, windowVariance, windowKurtosis, windowIQR];
         end
       end 
     end
@@ -209,29 +244,34 @@ end
 
 function value = getBinMean(data, idx, numBins)
 
+  #search for an idx == numBins+1. this index occurs if a value == lastEdge, thanks histc... 
+  isSix = find(idx == numBins+1);
+  if isSix != 0 
+    idx(isSix) = numBins;
+  endif
 
-value = [];
+  value = [];
   for i = 1:numBins
     flagBinMembers = (idx == i);
     binMembers = data(flagBinMembers);
     
-   # if length(binMembers) == 0
+    if length(binMembers) == 0
-     # idx
+      value(i) = 0;
-      #data
+    else
-     # input = 'balala'
+      value(i) = mean(binMembers);
-    #endif
+    endif
-    
-    value(i) = mean(binMembers);
   end  
 end
 
 #triangular functions. (edges of the triangles; num fft values -> nfft.)
 function triFilter = getTriangularFunction(edges, nfft)
 
+  global samplerateHZ;
+
   #get idx of the edges within the samples. thanks to fft each sample represents a frequency. 
   # idx * samplerate / nfft = hertz of that idx
   for i = 1:length(edges)
-    edgesByIdx(i) = floor((nfft + 1) * edges(i)/100); #100hz is the samplerate
+    edgesByIdx(i) = floor((nfft + 1) * edges(i)/samplerateHZ); 
   end
   
   #generate the triangle filters

toni/octave/genfeatures.m

@@ -1,4 +1,4 @@
-display("rawPlot")
+display("Generating Features")
 
 #load and plot raw data
 #{
@@ -14,9 +14,8 @@ access the cells using classes{u}.samples{v}.raw{w}
 #}
 source("functions.m");
 
-trainsetPerClass = 6; #number of used trainsets for one class
 classes = {};
-classes = getRawTrainData(trainsetPerClass);
+classes = getRawTrainData();
 
 #outPath = "/home/toni/Documents/handygames/HandyGames/toni/img/raw"
 #plotData(classes, outPath);
@@ -32,12 +31,12 @@ filteredClasses = filterData(classes);
 data structure of windowedClasses:
 classes[1 to 5]
   samples[1 to trainsetPerClass]
-    raw[1 to 9] <--- 9 different signals
+    raw[1 to 15] <--- 15 different signals
-      WindowsAccel[2.5s at 200ms] 
+      3x WindowsAccel (X, Y, Z)
         Win, Win, Win, Win ... <--- single matrices
-      WindowsGyro[2.5s at 200ms]
+      3x WindowsGyro (X, Y, Z)
         Win, Win, Win, Win ... <--- single matrices
-      WindowsMagnet[2.5s at 200ms]
+      3x WindowsMagnet (X, Y, Z)
         Win, Win, Win, Win ... <--- single matrices
         
     ---> add 6 additional sensors: pca and magnitude    
@@ -56,11 +55,13 @@ windowedClasses = windowData(filteredClasses);
 #calculated features for the 5 signales (x, y, z, MG, PCA) of a sensor
 #{
 data structure of features
-label | feature #1 | 2 | 3 ... 
+[classLabel, binMeans, rms, psd, windowMean, windowSTD, windowVariance, windowKurtosis, windowIQR]
-
 #}
 features = featureCalculation(windowedClasses);
 
-#train svm
+#save features
+save features.txt features;
+display("saved features into features.txt");
+
+
 
-#run svm

toni/octave/settings.m

@@ -0,0 +1,8 @@
+#global trainsetPerClass = 6; #number of used trainsets for one class
+global setDataDir = "/home/toni/Documents/handygames/HandyGames/daten/";
+global setWindowSize = 256; #in samples per window. even integer! 
+global setWindowSliding = 320; #in ms - (sampling rate is 10 ms.. so numSamples*10)
+global setAutocorrelationBinSize = 5;
+global setPSDBinSize = 10; 
+global samplerateHZ = 100;
+

toni/octave/todo.txt

@@ -3,6 +3,7 @@ daten mal plotten
 vorverarbeitung:
 low-pass-filter (-60dB at 20hZ)
 windowing (5s sliding at 200ms)
+Hinweis: die ersten 4 - 5 samples sind schon ziemlich mistig. kommen ganz gruselige werte. 
 
 segmentation (erstmal weglassen, da daten dafür nicht trainiert. bräuchten trainingsdaten von non-exercise die dann gelabeled sind)
 laufen ist z.b. keine übung! (man läuft ja zwischen übungen durch die gegend)

toni/octave/training.m

@@ -0,0 +1,49 @@
+#train features using svm
+display("Train Features")
+
+#load all features
+# features = [sampleLabel, classLabel, binMeans, rms, psd, windowMean, windowSTD, windowVariance, windowKurtosis, windowIQR];
+load "features.txt"; #matrix is also called features
+
+# split features into training and test features using leave-one-out method
+# class idx:
+# idx 1 -> forwardbend
+# idx 2 -> kneebend
+# idx 3 -> pushups
+# idx 4 -> situps
+# idx 5 -> jumpingjack
+
+# define which sampleSet is used as testset and not for training.
+leaveOut = find(features(:,1) == 3 & features(:,2) == 2); #sampleset 3 class 2
+testFeatures = features(leaveOut, :); #set testSignal
+features(leaveOut,:) = []; #remove the testFeatures
+features(:,1) = []; #remove the sampleLabel
+
+# bring the feature matrix into libsvm format. 
+# 1. the label vector: 
+trainLabel = features(:,1);
+
+# 2. sparse matrix with every feature in one column:
+features(:,1) = []; #remove the classLabel
+trainFeatures = sparse(features);
+
+# before training we need to scale the feature values
+minimums = min(trainFeatures);
+ranges = max(trainFeatures) - minimums;
+
+trainFeatures = (trainFeatures - repmat(minimums, size(trainFeatures, 1), 1)) ./ repmat(ranges, size(trainFeatures, 1), 1); 
+
+# training: svm with default settings
+model = svmtrain(trainLabel, trainFeatures);
+
+display("Classify Features")
+# for testing we need to scale again
+testLabel = testFeatures(:,2);
+testFeatures(:,1:2) = []; #remove the labels
+testFeatures = (testFeatures - repmat(minimums, size(testFeatures, 1), 1)) ./ repmat(ranges, size(testFeatures, 1), 1);
+
+# classification
+[predict_label, accuracy, dec_values] = svmpredict(testLabel, testFeatures, model);
+
+
+