void setup() {
stream.setLabelsForAllDimensions({"x", "y", "z"});
useInputStream(stream);
DTW dtw(false, true, null_rej);
dtw.enableTrimTrainingData(true, 0.1, 75);
pipeline.setClassifier(dtw);
pipeline.addPostProcessingModule(ClassLabelTimeoutFilter(timeout));
usePipeline(pipeline);
registerTuneable(
null_rej, 0.1, 5.0, "Variability",
"How different from the training data a new gesture can be and "
"still be considered the same gesture. The higher the number, the "
"more different it can be.",
[](double new_null_rej) {
pipeline.getClassifier()->setNullRejectionCoeff(new_null_rej);
pipeline.getClassifier()->recomputeNullRejectionThresholds();
});
registerTuneable(
timeout, 1, 3000, "Timeout",
"How long (in milliseconds) to wait after recognizing a "
"gesture before recognizing another one.",
[](double new_timeout) {
ClassLabelTimeoutFilter* filter =
dynamic_cast<ClassLabelTimeoutFilter*>(
pipeline.getPostProcessingModule(0));
assert(filter != nullptr);
filter->setTimeoutDuration(new_timeout);
});
}
int main (int argc, const char * argv[])
{
//Create a new gesture recognition pipeline
GestureRecognitionPipeline pipeline;
//Add an ANBC module
pipeline.setClassifier( ANBC() );
//Add a ClassLabelFilter as a post processing module with a minCount of 5 and a buffer size of 10
pipeline.addPostProcessingModule( ClassLabelFilter(5,10) );
//Load some training data to train and test the classifier
ClassificationData trainingData;
ClassificationData testData;
if( !trainingData.loadDatasetFromFile("ClassLabelFilterTrainingData.txt") ){
cout << "Failed to load training data!\n";
return EXIT_FAILURE;
}
if( !testData.loadDatasetFromFile("ClassLabelFilterTestData.txt") ){
cout << "Failed to load training data!\n";
return EXIT_FAILURE;
}
//Train the classifier
if( !pipeline.train( trainingData ) ){
cout << "Failed to train classifier!\n";
return EXIT_FAILURE;
}
//Use the test dataset to demonstrate the output of the ClassLabelFilter
for(UINT i=0; i<testData.getNumSamples(); i++){
VectorDouble inputVector = testData[i].getSample();
if( !pipeline.predict( inputVector ) ){
cout << "Failed to perform prediction for test sampel: " << i <<"\n";
return EXIT_FAILURE;
}
//Get the predicted class label (this will be the processed class label)
UINT predictedClassLabel = pipeline.getPredictedClassLabel();
//Get the unprocessed class label (i.e. the direct output of the classifier)
UINT unprocessedClassLabel = pipeline.getUnProcessedPredictedClassLabel();
//Also print the results to the screen
cout << "Processed Class Label: \t" << predictedClassLabel << "\tUnprocessed Class Label: \t" << unprocessedClassLabel << endl;
}
return EXIT_SUCCESS;
}
bool GRT_Recognizer::initPipeline(string trainingdatafile, int dimension)
{
//Initialize the training and info variables
// infoText = "";
// trainingClassLabel = 1;
// noOfHands = 2;
//noOfTrackedHands = 0;
//The input to the training data will be the R[x y z]L[x y z] from the left end right hand
// so we set the number of dimensions to 6
LabelledTimeSeriesClassificationData trainingData;
//trainingData.setNumDimensions(6);
trainingData.loadDatasetFromFile(trainingdatafile);
//Initialize the DTW classifier
DTW dtw;
//Turn on null rejection, this lets the classifier output the predicted class label of 0 when the likelihood of a gesture is low
dtw.enableNullRejection( true);
//Set the null rejection coefficient to 3, this controls the thresholds for the automatic null rejection
//You can increase this value if you find that your real-time gestures are not being recognized
//If you are getting too many false positives then you should decrease this value
dtw.setNullRejectionCoeff(2);
//Turn on the automatic data triming, this will remove any sections of none movement from the start and end of the training samples
dtw.enableTrimTrainingData(true, 0.1, 90);
//Offset the timeseries data by the first sample, this makes your gestures (more) invariant to the location the gesture is performed
dtw.setOffsetTimeseriesUsingFirstSample(true);
//Add the classifier to the pipeline (after we do this, we don't need the DTW classifier anymore)
pipeline.setClassifier( dtw );
//pipeline.addPreProcessingModule(MovingAverageFilter(5,dimension));
//pipeline.addFeatureExtractionModule(FFT(16,1, dimension));
/*ClassLabelFilter myFilter = ClassLabelFilter();
myFilter.setBufferSize(6);
myFilter.setBufferSize(2);*/
pipeline.addPostProcessingModule(ClassLabelChangeFilter());
pipeline.train(trainingData);
return true;
}
void setup() {
stream.setLabelsForAllDimensions({"audio"});
pipeline.addFeatureExtractionModule(
FFT(kFftWindowSize, kFftHopSize,
DIM, FFT::HAMMING_WINDOW, true, false));
MFCC::Options options;
options.sample_rate = kSampleRate;
options.fft_size = kFftWindowSize / 2;
options.start_freq = 300;
options.end_freq = 3700;
options.num_tri_filter = 26;
options.num_cepstral_coeff = 12;
options.lifter_param = 22;
options.use_vad = true;
options.noise_level = noise_level;
pipeline.addFeatureExtractionModule(MFCC(options));
pipeline.setClassifier(SVM());
// GMM(16, true, false, 1, 100, 0.001));
// In post processing, we wait #n predicitons. If m out of n predictions are
// from the same class, we declare the class as the right one.
//
// n = (duration * sample_rate) / frame_size
// where duration = post_duration
// sample_rate = kSampleRate
// frame_size = kFftHopSize
// m = n * post_ratio
int num_predictions = post_duration / 1000 * kSampleRate / kFftHopSize;
pipeline.addPostProcessingModule(
ClassLabelFilter(num_predictions * post_ratio, num_predictions));
auto ratio_updater = [](double new_ratio) {
ClassLabelFilter* filter =
dynamic_cast<ClassLabelFilter*>(pipeline.getPostProcessingModule(0));
// Recalculate num_predictions as post_duration might have been changed
int num_predictions = post_duration / 1000 * kSampleRate / kFftHopSize;
filter->setMinimumCount(new_ratio * num_predictions);
};
auto duration_updater = [](int new_duration) {
ClassLabelFilter* filter =
dynamic_cast<ClassLabelFilter*>(pipeline.getPostProcessingModule(0));
// Recalculate num_predictions as post_duration might have been changed
int num_predictions = post_duration / 1000 * kSampleRate / kFftHopSize;
filter->setBufferSize(num_predictions);
};
auto noise_updater = [](int new_noise_level) {
MFCC *mfcc = dynamic_cast<MFCC*>(pipeline.getFeatureExtractionModule(1));
mfcc->setNoiseLevel(new_noise_level);
};
registerTuneable(noise_level, 0, 20,
"Noise Level",
"The threshold for the system to distinguish between "
"ambient noise and speech/sound",
noise_updater);
registerTuneable(post_duration, 0, 2000,
"Duration",
"Time (in ms) that is considered as a whole "
"for smoothing the prediction",
duration_updater);
registerTuneable(post_ratio, 0.0f, 1.0f,
"Ratio",
"The portion of time in duration that "
"should be from the same class",
ratio_updater);
useInputStream(stream);
useOutputStream(oStream);
usePipeline(pipeline);
useLeaveOneOutScoring(false);
setGUIBufferSize(kSampleRate);
}
