Date of Award
Daniel Felix Ritchie School of Engineering and Computer Science, Mechanical and Materials Engineering
Peter J. Laz
Casey A. Myers
Body model, Computer vision, Optimization, Point cloud
While marker-based motion capture remains the gold standard in measuring human movement, accuracy is influenced by soft-tissue artifacts, particularly for subjects with high body mass index (BMI) where markers are not placed close to the underlying bone. Obesity influences joint loads and motion patterns, and BMI may not be sufficient to capture the distribution of a subject’s weight or to differentiate differences between subjects. Subjects in need of a joint replacement are more likely to have mobility issues or pain, which prevents exercise. Obesity also increases the likelihood of needing a total joint replacement. Accurate movement data for subjects with a higher BMI is of the utmost importance because it can be used to inform treatment options for people who have received joint replacements or are waiting to receive a replacement. Currently, movement data from this subgroup of people has error introduced due to soft tissue artifacts from markered motion capture, such as Vicon. By investigating ways of measuring movement with computer vision tools, the influence of soft tissue artifacts can be investigated further in movement data. The Azure Kinect DK is a depth camera that collects point cloud data of the surface of the person and automatically calculates joint centers. The objectives of this thesis were to 1) Design a fast and accurate procedure to generate a full-body point cloud with two Azure Kinect DK cameras; 2) Create subject-specific computational representations by fitting the Skinned Multi-Person Linear (SMPL) model to Kinect point cloud data. With IRB approval, 24 subjects consented to perform T-pose , lunge, sit to stand, and walking activities while being recorded by two Azure Kinect DK depth cameras. For all activities except walking, one camera was facing the subject head on and the other recorded the right sagittal view. During walking, the subjects walked toward both cameras. Of these 24, point cloud and joint data from 16 subjects were used. This study uses exclusively data from the T-pose trial.
Without post-processing, these two views of the T-pose do not produce complete data of the surface of a person. A technique was developed to recreate missing sections of the body and create a partial point cloud. Another method of collecting point cloud data was also explored using two cameras and a turn table. A synchronous capture is taken of the front and side, the person is rotated 180 degrees, and another capture is taken. An optimization pipeline was designed to fit the SMPL model to both types of data. The process created high-quality computational representations of each subject. During data collection, extensive anthropometric measures were recorded of each subject with a tailor’s tape and used to verify the accuracy of the model by comparing them to digitally recreated measures on each SMPL model. The validity of the model was quantified with a percent error calculation between subject manual and SMPL measurements. Subject manual and SMPL measurements are highly correlated with an R2 > 0.9 and p-value << 0.1. Across all measurements and subjects, there is an average absolute percent error of 4.71 ± 4.09% The average absolute percent error between any measurements never exceeds 10%. The largest absolute percent error is in the ankle with 9.80 ± 6.33% , and the smallest absolute percent error is in the floor to shoulder measurement with 1.17 ± 0.76%.
This tool to create a subject-specific whole-body computational representation has a broad variety of applications. This model fitting process is ready to be deployed on the point cloud collection processes described and will be useful in creating subject-specific finite element models. The Azure Kinect DK gives the ability to perform markerless motion capture and surface point cloud collection in a clinical setting, such as a doctor’s office. These body models could be generated from data like this and allow for patient classification. The SMPL model’s ability to describe realistic body shapes could be used to accurately calculate the moment of inertia for many body segments. A standardized tool to measure body shape and habitus can allow for the discovery of correlations between body dimensions and movement patterns.
Copyright Statement / License for Reuse
All Rights Reserved.
Copyright is held by the author. User is responsible for all copyright compliance.
Received from ProQuest
Young, Emma, "Novel Approach for Non-Invasive Prediction of Body Shape and Habitus" (2023). Electronic Theses and Dissertations. 2330.
Engineering, Computer science