Measuring the motion of human hand joints is a challenging task due to the high number of DOFs. In this study, we proposed a low-cost hand tracking system built on action cameras and ArUco markers to measure finger joint rotation angles. The lens distortion of each camera was corrected first via intra-calibration and the videos of different cameras were aligned to the reference camera using a dynamic time warping based method. Two methods were proposed and implemented for extracting the rotation angles of finger joints: one is based on the 3D positions of the markers via inter-calibration between cameras, named pos-based method; the other one is based on the relative marker orientation information from individual cameras, named rot-based method. An experiment was conducted to evaluate the effectiveness of the proposed system. The right hand of a volunteer was included in this practical study, where the movement of the fingers was recorded and the finger rotation angles were calculated with the two proposed methods, respectively. The results indicated that although using the rot-based method may collect less data than using the pos-based method, it was more stable and reliable. Therefore, the rot-based method is recommended for measuring finger joint rotation in practical setups.

A high-fidelity digital representation of (part of) the human body is a key enabler for integrating humans in a digital twin. Among different parts of human body, building the model of the hand can be a challenging task due to the posture deviations among collected scans. In this article, we proposed a posture invariant statistical shape model (SSM) of the human hand based on 59 3D scans of human hands. First, the 3D scans were spatially aligned using a Möbius sphere-based algorithm. An articulated skeleton, which contains 20 bone segments and 16 joints, was embedded for each 3D scan. Then, all scans were aligned to the same posture using the skeleton and the linear blend skinning (LBS) algorithm. Three methods, i.e., principal component analysis (PCA), kernel-PCA (KPCA) with different kernel functions, and independent component analysis (ICA), were evaluated in the construction of the SSMs regarding the compactness, the generalization ability, and the specificity. The PCA-based SSM was selected, where 20 principal components were used as parameters for the model. Results of the leave-one-out validation indicate that the proposed model was able to fit a given 3D scan of the human hand at an accuracy of 1.21 ± 0.14 mm. Experiment results also indicated that the proposed SSM outperforms the SSM that was built on the scans without posture correction. It is concluded that the proposed posture correction approach can effectively improve the accuracy of the hand SSM and therefore enables its wide usage in human-integrated digital twin applications.