The Previous systems for 3D human hand posture estimation have adopted clustered multi-layer large-scale database with narrowing of search space by its past estimation results. But once an estimated result at a time is out of the search space, the system can't find out a true or optimal value. Our system therefore has adopted non-clustered large-scale database with narrowing of search space, rather, a coarse search at the first stage according to some aspects of inputted hand images, and an accurate search at the second stage with low-order image features. The experimental results showed that the averaged estimation error is -2.11 degrees, and the candidates for accurate search at the second stage are reduced from 28,386 to 137.7 data sets, which indicates our system realizes the stable hand posture estimation with high accuracy and processing speed as previous system without using the past results.
