Hand Pose Estimation Based On Point Cloud

As a key technology in human-computer interaction,hand pose estimation is widely used in various fields,especially in some applications of virtual reality and augmented reality,which can improve the user experience and realize the direct interaction between human hand and virtual object.In recent years,with the rapid development of artificial intelligence technology,the current three-dimensional data acquisition equipment has also been rapid development.However,due to the high dimension of 3D hand pose,the great change of hand direction,and the high self phase and self occlusion between fingers,the estimation results have the problems of low accuracy and poor robustness.Therefore,in this paper,the depth map is transformed into a point cloud representation as the input,and the point cloud depth neural network is constructed.Through this way,the point cloud features are learned for the task of 3D hand pose estimation.Through this method,the 3D information in the depth map can be fully utilized,and the data is too jumbled.Firstly,by studying the structure and principle of pointnet and hierarchical pointnet,this paper uses a hand pose estimation method based on hierarchical pointnet.Firstly,the data set and the process of data preprocessing are introduced.This paper focuses on the hand region locking and the hierarchical pointnet network with three feature extraction layers.Secondly,in order to further improve the accuracy of hand pose estimation,this paper uses a new network composed of hierarchical pointnet network and point attention converter to do the task of hand pose estimation.Therefore,this paper proposes a hand pose estimation method combined with point attention converter.Due to the independence and correlation between the joints of the hand,the function of the point attention converter is to take the relationship between the joint points as the feature input,so that the accuracy of hand attitude estimation can be further improved.At the same time,Gumbel sampling method(GSS)is used to replace the farthest point sampling(FPS)for down sampling.Finally,experiments are carried out with the above two methods and compared with other hand pose estimation methods.This paper is published in cvpr2015?The experimental results show that the hand pose estimation method based on hierarchical pointnet network is better than most methods in estimation accuracy,while the hand pose estimation method based on point attention converter is better than the former in the final average estimation accuracy,However,the accuracy of joint point estimation in fingertip position is slightly lower than that based on hierarchical pointnet.