Efficient 3D Hand Posture Estimation with Self-Occlusion from Multiview Images

Automatic initialization of human hand posture is an important task in human-machine interaction systems using hand gesture. We present a part-based 3D hand posture estimation approach which combines top-down with bottom-up approaches. We use a factor graphical structure model to inference hand pose with self-occlusions. The location probability distribution of each part is determined by evaluating appropriate likelihood. To reduce a heavy computation burden caused by enormous size of each part state spaces, we use genetic algorithm to prune the low likelihood state of each part. Our experimental evaluation demonstrates that improvements of inference optimization efficiency can be obtained with the proposed approach.