3차원 손 형상 모델과 인체 참조점을 이용한 손가락 관절 회전 중심 추정

Abstract

In digital human modeling, a finger joint center of rotation (CoR) which refers to the center of relative rotation between two adjacent bone segments needs to be estimated for shaping the link model. Joint CoR can be divided into two types: (1) fixed joint CoR, which has a fixed position regardless of hand posture, and (2) instantaneous joint CoR, whose position changes according to hand posture. Joint CoRs can be estimated from medical scan data (skeleton-based estimation) or motion capture data (surface-based estimation). The skeleton-based estimation provides accurate joint CoR compared to surface-based estimation. However, there’s a risk of exposure to radiation due to CT/MRI scanning. In addition, the process of aligning the bone segments takes a lot of time. On the other hand, the surface-based estimation is safe and fast in data acquisition. However, the joint CoR may not be estimated as accurate as the skeleton-based estimation because of the soft tissue deformation. The present study is intended to construct 3D hand models and develop regression models to estimate fixed and instantaneous joint CoRs using surface landmarks. The proposed study consists of three steps: (1) constructing 3D hand models including bone shape, fixed joint CoR location, instantaneous joint CoR location, and surface mesh, (2) developing regression models that can estimate the coordinates of joint CoR based on the coordinates of surface landmarks, (3) evaluating the performance of the model through the mean distance (MD) between the joint CoRs estimated through the developed regression model and the constructed reference joint CoRs. First, 3D hand models for 9 participants (male 6, female 3) with 10 postures per person were constructed through the CT scanning experiment. The bone shape of the template posture (posture 1) obtained from the CT scan image was divided into 29 bone segments and registered to the rest of the posture (posture 2 ~ 10). The joint CoR data were constructed for all the postures (posture 1 ~ 10) using joint CoR data estimated from the skeleton-based method by Lim et al. (2018). The surface mesh was obtained with the bone shape from the CT scan image and the bone was removed using RapidForm2006 (Inus Technology, Inc., Korea) which is 3D image processing software. The template mesh was aligned with the remaining target meshes through template registration and the number and positions of vertices of the surface mesh were standardized. Second, a novel regression model for joint CoR estimation was established using coordinates of 4 surface landmarks around the joint CoRs. Twelve joint CoRs (distal interphalangeal (DIP), proximal interphalangeal (PIP), and metacarpophalangeal (MCP) of the index finger, middle finger, ring finger, and little finger) were estimated. Three landmark sets (registered surrounding set, registered dorsal set, and manual dorsal set) were established according to the landmark locations (surrounding, dorsal), and landmarking method (manual, registration). The landmarks of the surrounding set are located in the dorsal, basal, radial and ulnar directions of each joint CoR, whereas the landmarks of the dorsal set are located on the dorsum of the finger to which the joint CoR belongs. Unlike manual landmarking, in which landmarks are manually inserted into the surface mesh by an analyst, the registration landmarking is a method of inserting landmarks into the template mesh only and registering the landmarks on the target mesh by template registration. The models for estimating fixed joint CoR and instantaneous joint CoR were developed separately. Third, the constructed hand models were divided into four sets and cross validation was performed to evaluate the performance of the established regression model. In the test set, the distances between the estimated joint CoRs and the reference joint CoRs were calculated, and the estimation performance of the regression model was evaluated based on the mean distance (MD). In both the fixed and instantaneous joint CoR estimation models, the MD of the registered surrounding set and that of the registered dorsal set showed no significant difference for 10 out of 12 joint CoRs. The manual dorsal set showed significantly lower MD for 8 out of 12 joint CoRs than the registered dorsal set. There was no significant difference between the fixed joint CoR estimation model and the instantaneous joint CoR estimation model except for the MCP of the little finger when the landmark set was same. The surface landmark-based fixed joint CoR estimation model developed in this study has improved overall performance compared to the surface-based fixed joint CoR estimation model of Lim et al. (2018). Especially, the MD of the manual dorsal landmark set-based estimation model was found to be 24.8% ~ 58.4% lower in DIP, 35.5% ~ 60.4% lower in PIP, and 43.3% ~ 73.1% lower in MCP compared to that of the estimation model of Lim et al. (2018).