Research On 3D Hand Pose Estimation Based On Single RGB Image

With the unceasing progress in the computer vision field and increasing demand for humancomputer interaction,it is noticeable that the estimation approach of hand pose which based on computer vision has emerged constantly in recent years.Not only the appearance of deep convolution network has promoted hand pose estimation study advancing,but also mirrors the significant application and shows evidence of remarkable research potential.There are two leading three-dimensional hand pose estimation methods so far: the one is based on ordinary monocular RGB images;the other depends on depth images and RGB-D sequences.In several hand pose estimation approaches that on account of depth images,some of them have already achieved positive experimental results,but its facilities,for instance,the depth camera,are too expensive to popularize;on the other hand,the depth camera has to operate in a stable interior space.Thus such an obstacle restricts the implementation of hand pose estimation methods based on depth images.The mission of hand pose estimation based on RGB images still faces challenges.Firstly,it is hard to take all image feature of two-dimension information due to the high flexibility of hands,which result in the inaccuracy of the three-dimensional estimation;secondly,the imagery of the hand is too small to capture in a full-size picture,so that affects the hand gesture in two-dimensional and three-dimensional estimative accuracy.Therefore,this paper proposes a three-dimensional hand pose estimation network based on RGB images,which effectively improves the above-mentioned problems.Aiming at the problem of small scale hand imagery,this paper designs and implements a lightweight cropping method,as well as utilize the principle of morphological expansion to improve the cropping effect.The position of hands in the full-size image can be recognized precisely by cropping the original image so as to improve the revolution of the hand part,and the accuracy of hand pose estimation thus increased at the same time.In addition,subsequent 3D hand pose estimation experimentation,the essay proposes and implements a method for restoring 3D hand pose based on the 2D heat map.Compared to the threedimensional hand pose estimation method,the accuracy has prominently improved.In order to solve the issues of hand flexibility and the transformation from 2D to 3D message directly,this paper designs and implements a dual-channel network.The 2D results will be upgraded directly to 3D posture by using an attention mechanism structure.Then the standard coordinate system establishment is needed,using the perspective estimation network to observe the radiation relationship between the standard coordinate system and the original coordinate system,so as to optimize the accuracy of the three-dimensional posture estimation network.Accordingly,experiments have proved that the method effectively improves the accuracy of 3D hand pose results indeed.