Head Tracking in Meeting Scenarios Sascha Schreiber
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 2/18 Overview Video processing system for person action recognition Single person head tracking using elliptical structures Multiple person tracking Summary and outlook
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 3/18 System Setup 1)Face/Head Tracking: Face Detection & Particle Filtering 2)Feature extraction: Global Motion Features 3)Kalman filtering: necessary for disturbed feature streams 4)Temporal segmentation: Bayesian Information Criterion 5)HMM classification: 6 models, one for each gesture Video streamTracked personsKalman filteringTemporal segmentationHMM classification Standing up Sitting down Pointing Nodding Shaking head Writing Gesture inventory:
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 4/18 Single Person Tracking Former system: –Initialization by pyramid sampling slow, only frontal to half-profile views of faces can be detected –Weighting of particles by the output of a neural network Loss of track, if no face visible (e.g. person leans forward) Particles have to be placed quite exactly
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 5/18 Actual system: –Initialization by skin colored regions Single Person Tracking Former system: –Initialization by pyramid sampling slow, only frontal views of faces can be detected 1) Skin color blob detection 2) Search for valid blobs ratio major axis/minor axis
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 6/18 Former system: –Weighting of particles by the output of a neural network Loss of track, if no face visible (e.g. person leans forward) Particles have to be placed quite exactly Actual system –Weighting of particles using ellipses Single Person Tracking 1) Compute gradient image 2) Search best fitting gradient along normal: 3) Calculate cost function:
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 7/18 Single Person Tracking Comparision of tracking results: Neural netEllipses
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 8/18 Multiple Person Tracking Problem: -Particles concentrate on location with highest head likelihood Solution: -Introduction of „Super particles“ representing the number of persons in scenario
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 9/18 Multiple Person Tracking Weighting of Superparticles 1) Counting blobs in binary skin color mask n blobs 2) Testing all possible combinations
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 10/18 Multiple Person Tracking Weighting of Superparticles Configuration coverage Configuration compactness Particle likelihood 1) Counting blobs in binary skin color mask n blobs 2) Testing all possible combinations 3) Measurement / Cost function
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 11/18 Multiple Person Tracking Weighting of Superparticles 1) Counting blobs in binary skin color mask n blobs 2) Testing all possible combinations 3) Measurement / Cost function
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 12/18 Multiple Person Tracking Demovideo „Multiple Person Tracking“
24/09/2004 Institute for Human-Machine Communication Munich University of Technology Sascha Schreiber 13/18 Summarization & Outlook Robust head tracking Improving tracking algorithm Features computed relative to the head position Tracking hands as additional feature Temporal segmentation facilitated by tracking Combined tracking and recognition
Head Tracking in Meeting Scenarios Sascha Schreiber