Welcome to Journal of Graphics share: 

Journal of Graphics ›› 2022, Vol. 43 ›› Issue (5): 892-900.DOI: 10.11996/JG.j.2095-302X.2022050892

• Computer Graphics and Virtual Reality • Previous Articles     Next Articles

A 3D hand pose estimation method based on improved PointNet++ 

  

  1. School of Information, North China University of Technology, Beijing 100144, China
  • Online:2022-10-31 Published:2022-10-28
  • Supported by:
    General Project of Beijing Science Foundation (18YTC038); Youth Fund Project of Beijing Natural Science Foundation (4194076); General Project of Scientific Research Plan of Beijing Municipal Commission of Education (KM201910009014) 

Abstract:

To address the problem that the processing of local features of point cloud in PointNet++ network sometimes results in a large amount of computation due to the large grouping range, a 3D hand pose estimation method based on the improved PointNet++ network was proposed. Firstly, the gesture point cloud was triangulated based on the combination of Delaunay triangulation algorithm and K-Median clustering algorithm, thus creating the triangular mesh model of the gesture point cloud. Simultaneously the average edge length of the triangular mesh model was calculated. Then, with the average edge length of the triangular mesh model as the radius, the points sampled by the farthest point sampling (FPS) algorithm were searched by ball query. Then the sampled point cloud was grouped by K-Nearest Neighbors algorithm according to the maximum value of the searched sampled points. Finally, the grouped point cloud was input into the PointNet to perform the 3D hand pose estimation. The improved PointNet++ network can automatically adjust the number of local abstraction points of point cloud grouping according to point cloud density at different levels. Experiments show that, without affecting the accuracy of 3D hand pose estimation, the proposed method can enhance the training speed of PointNet++, as well as effectively reducing the  computation of feature extraction in 3D hand pose estimation, so that the computer can capture the hand pose more quickly. 

Key words:  , 3D hand pose estimation, PointNet++, Delaunay triangulation, ball query search, K-nearest neighbor search 

CLC Number: