Human Hand Pose Estimation Based On Point Cloud

With the advent of artificial intelligence boom,human-computer interaction technology has been greatly developed,and has a large number of applications in the fields of network live broadcast,game interaction,and augmented reality.Among them,human-hand interaction plays the most important role in human-computer interaction.The human hand interaction problem is based on human hand detection,human hand pose estimation and human hand semantic understanding as the technical support.Accurate human hand pose estimation is the most critical part of the human hand interaction problem,and it has a connecting role.The problem of human hand posture estimation is essentially a regression problem,which returns to the position of the human hand joint point in three-dimensional space through a picture of the hand.The difficulty of this problem is: first,the human hand is a non-rigid object and has a high degree of freedom morphological structure;second,the human hand is extremely prone to self-occlusion during imaging,resulting in the lack of extracted picture information.At the same time,in view of the real-time and mobility characteristics of the interaction problem,the feature extraction of the human hand poses the requirements of faster estimation speed and higher regression accuracy.The point cloud-based pose estimation technology has the advantages of smaller models and higher accuracy than depth images.Based on the point cloud data,the main research contents of this paper are as follows:First,design a two-stage cascade pose estimation network.In the first stage,the local information extracted from the original information is fused into global information and the initial joint point position is obtained.Then,in the second stage,the initial joint point position is input into the network.The global information is used to extract the local information again to optimize the joint position.Second,inspired by graph theory,a graph convolutional neural network that is more suitable for point cloud data is used for feature extraction.This paper proposes for the first time structured human features as the output and uses the topological relationship to constrain the human hand to improve the efficiency of feature utilization.After sufficient experimental verification,the two algorithms proposed in this paper can achieve advanced algorithm performance under a variety of evaluation indicators.