一种基于改进PointNet++网络的三维手姿估计方法

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

针对PointNet++网络处理点云局部特征时因分组范围区过大导致计算量较大的问题,提出一种改进的PointNet++网络的三维手姿估计方法。首先对手势点云进行基于Delaunay三角剖分算法与K中位数聚类算法相结合的三角剖分,得到手势点云的三角网格模型,并计算三角网格模型的边长均值;然后以三角网格模型边长均值为半径,对最远点采样(FPS)的采样点进行球查询搜索,再根据搜索到的采样点个数极值对采样点云进行K近邻分组,并最终输入PointNet网络,完成三维手姿的位置估计。改进后的PointNet++网络可以根据不同的点云密度自动调整网络分组区域的局部提取点个数。实验结果表明,在不影响三维手姿估计精度的情况下,该方法提高了PointNet++网络的模型训练速度,并在三维手姿估计中可有效减少特征提取的计算量,使计算机能够更快地捕捉手姿状态。

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.