End-to-end training of convolutional neural network for 3D hand pose estimation in dual-view RGB image

Abstract

Gesture based robot human-computer interaction is more natural and convenient. Accurate hand pose estimation has always been a research hotspot in the field of human-computer interaction. This paper proposes an end-to-end training hand pose estimation method based on convolutional neural network for ordinary RGB gesture images. This method can establish a data set containing pose data for the hand to be estimated, and use convolutional neural network to fit gesture images and pose angle data. In addition, aiming at the problem that gesture self-occlusion affects the accuracy of hand pose estimation, this paper proposes a strategy of hand pose estimation model training and prediction using dual-view images. Through experimental verification and analysis, the hand pose estimation method has achieved good test results on the unknown test sample set. The method has the advantages of simple design, convenient use and high accuracy.