/**
* First quantizes the feature vector against codebook, and then runs viterbi decoding
*
* @param rf
* @return
*/
public String recognizeGesture(RawFeature rf) {
codebook = new Codebook();
GestureFeature[] gestureFeatures = getFeature(rf);
Points[] pts = getPointsFromFeatureVector(gestureFeatures);
int[] quantized = codebook.quantize(pts);
String[] regGestures = readRegGestureModels();
HiddenMarkov[] hmms = new HiddenMarkov[regGestures.length];
// read hmms
for (int i = 0; i < hmms.length; i++) {
hmms[i] = new HiddenMarkov(regGestures[i]);
}
// find likelihood by viterbi decoding of quantized seq
double[] likelihoods = new double[regGestures.length];
for (int j = 0; j < likelihoods.length; j++) {
likelihoods[j] = hmms[j].viterbi(quantized);
}
// find the largest likelihood
double highest = Double.NEGATIVE_INFINITY;
int wordIndex = -1;
for (int j = 0; j < regGestures.length; j++) {
if (likelihoods[j] > highest) {
highest = likelihoods[j];
wordIndex = j;
}
}
// best matched
return regGestures[wordIndex];
}
/**
* Performs hmm train for all training data set.<br>
* First quantizes the feature vector against codebook, and then train baum-welch
*
* @return
*/
public boolean trainHMM() {
boolean operationSuccess = false;
codebook = new Codebook();
database.setMode(DBMode.TRAINDATA);
Model[][] regModels = database.readAllDataofCurrentMode();
String[] gestName = database.getRegisteredModelNames();
int quantizedSeq[][];
HiddenMarkov mkv = new HiddenMarkov(NUM_STATES, NUM_SYMBOLS);
// for each gesture
for (int i = 0; i < regModels.length; i++) {
operationSuccess = false;
// for each train sample of current gesture
quantizedSeq = new int[regModels[i].length][];
for (int j = 0; j < regModels[i].length; j++) {
GestureFeature[] gf = getFeature((RawFeature) regModels[i][j]);
Points[] pts = getPointsFromFeatureVector(gf);
quantizedSeq[j] = codebook.quantize(pts);
}
mkv.setTrainSeq(quantizedSeq);
mkv.train();
mkv.save(gestName[i]);
operationSuccess = true;
}
return operationSuccess;
}