Articulated Hand Tracking by ICA-based Hand Model and Multiple Cameras

Abstract

In this chapter, we proposed three new approaches, the ICA-based hand model, articulated hand motion tracking by multiple cameras, and Particle filtering with prediction. The ICA-based hand model is the ICA-based representation of hand articulation for tracking hand-finger gestures in image sequences. The dimensionality of the hand motion space is reduced by PCA and then ICA is applied to extract the local feature vectors. In the ICAbased model, each of the first five basis vectors corresponds to a particular finger motion, because the joints in each finger have stronger dependencies than the joints across different fingers. In the ICA-based model, hand poses can be represented by five parameters with each parameter corresponding to a particular finger motion. We implemented articulated hand motion tracking by particle filter using this ICA-based hand model. Experimental results show that the ICA-based model is very useful for articulated hand tracking in image sequences. Next approach is an articulated hand motion tracking by multiple cameras. This method is useful for gesture recognition. Tracking a free hand motion against a cluttered background was unachievable in previous methods because hand fingers are self-occluding. To improve search efficiency, we proposed adding prediction to particle filtering so that more particles are generated in areas of higher likelihood. The experimental results show that our method can correctly and efficiently track the hand motion throughout the image sequences even if hand motion has large rotation against a camera. The methods in this chapter are easily extended to many other visual motion capturing tasks.