High-speed search for similar image from database for dexterous robot hand control

The authors propose a system for searching the posture of human 3D hands and fingers in real time and with high accuracy, without using any special peripheral equipment such as infrared range sensor, PC cluster, etc., by a method of searching similar image quickly with high accuracy from a large volume of image database containing complicated shapes and self-occlusions. In designing the system, we constructed a database to be adaptable even to differences in hand shape among individuals, and searched computer graphics (CG) images of hand similar to unknown hand image input, through extraction of characteristics using high-order local autocorrelational patterns, reduction of the amount of characteristics centering on principal component analysis, and prior rearrangement of data corresponding to the amount of characteristics. As a result of experiments, our system carried out high-accuracy estimation of human hand shape with the processing speed of 30 fps or over