Estimation of distance between thumb and forefinger from hand dorsal image using deep learning
Takuma Shimizume, Takeshi Umezawa, Noritaka Osawa · 2018
A three-dimensional virtual object can be manipulated by hand and finger movements with an optical hand tracking device where it is necessary to recognize a posture of one's hand. Conventional hand posture recognition is based on three-dimensional coordinates of fingertips and a skeletal model of the hand [1]. It is difficult for conventional methods to estimate a posture of the hand when a fingertip is hidden from an optical camera. This study, therefore, proposes the estimation of a posture of a hand on the basis of a hand-dorsal image that can be taken even when the hand occludes its fingertips. A regression model that estimates a distance between fingertips of the thumb and forefinger was constructed using a convolution neural network (CNN) [2]. This work evaluated the root mean squared error (RMSE) of estimation. The RMSE of estimation based on a model on the same day was less than 1.8 mm, which shows that the proposed method could be an effective method where self-occlusion is a problem. This study also evaluates the robustness of the learning model to time-variation.