Research On Hand Pose And Shape Estimation Method Based On Single Color Image

Hand pose estimation has broad application potential in many intelligent fields,and the ways to obtain color images are rich and convenient.It is meaningful to estimate hand pose from a single color image.However,due the depth ambiguity,inferring the 3D pose directly from a single color image leads to a poor estimation results.In addition,only estimating the coordinates of abstract 3D hand keypoints cannot meet the needs of current practical applications.Therefore,it is necessary to conduct in-depth research on the reconstruction of reasonable and realistic hand shape.In view of the above problems,this paper conducts in-depth study,and the specific work is as followed:(1)Aiming at the problem of the depth ambiguity,this paper gives a weak depth information representation,which divides the depth range of hand into several depth intervals,so that each hand keypoint belongs to a specific depth interval and has a specific depth interval code.Then,the depth interval codes of the entire hand is combined in a certain order to generate the depth relation vector that can reveal the structural relationship of the hand in the depth space.The depth relation vector is used to generate 3D pose intermediate feature containing local hand keypoint position information and global hand structure information in the depth space.This feature is between 2D pose and 3D pose,and can be used as an intermediate transition in the 3D pose estimation process,so as to realize the mapping from the 3D pose intermediate feature to the 3D pose.This method shows that finding an appropriate intermediate transition in the process of 3D pose estimation is helpful for 3D pose estimation.(2)Considering that the simultaneous estimation of 3D pose and shape is more conducive to provide richer and clearer hand information,this paper gives 3D hand shape and pose estimation method based on MANO model.Without supervising the 3D hand shape label,this method can also generate reasonable and realistic hand shape,and at the same time obtain a relatively accurate 3D pose that conforms to physiological structure.This method uses the estimated relatively accurate 3D hand pose as the initial pose to reconstruct the hand shape,and realizes the 3D shape reconstruction of inverse kinematics.The reconstructed 3D shape passes through a specific keypoint regression and adjustment layer to obtain the 3D hand pose corrected by physiological structure.The computer simulation experiment results show the effectiveness of the two methods given in this paper.