Due to the difficulty of capturing complex structure of hands, 3D hand pose estimation in single depth image still suffers from the issues of low accuracy and poor robustness. In order to solve these problems, this paper proposed a 3D hand pose estimation method based on improved PointNet. The method firstly used a bounding box localization network to predict a 3D bounding box, and thereby accurately cropped the hand region. Then, it represented the depth image of the hand with point cloud. The point cloud that used to model the visible surface of the hand, could effectively utilize the 3D information in the depth image. At last, the hand point cloud was input into the improved PointNet to accurately estimate 3D hand pose. By introducing a jump connection, the improved PointNet made full use of the features of different levels, so it could capture the complex structure of the hand. The experimental results on NYU hand pose dataset show that the proposed method outperforms most of the existing methods, and the network is simple in structure, easy to train, and fast to run.

[2020-10-05] Printed Article /doc/2019.06.0171/v2