Parametric Modeling of Human Body Shapes using Principal Component Analysis

Abstract

The realistic and exact human body shape models are powerful tools used in various fields including ergonomics, garment design, biomechanics, and computer graphics. In ergonomics and garment design, such models are utilized for interaction analysis to be accurate and reliable, between users and products considering human factors. This ultimately enables better human-centered design of human spaces (e.g., car compartments, clothing, and manufacturing). In the biomechanics field, for example, these models are utilized in crash simulations essential for the occupant safety design of vehicles. The ability to generate detailed human body shapes enables the creation of virtual humans, including human bones or internal organs (e.g. liver, stomach, lungs, etc.), and consequently more accurate biomechanical responses can be predicted through crash simulations. With respect to computer graphics, these models are used to generate very realistic computer animations of human characters. This thesis proposes a new statistical modeling framework based on principal component analysis for parametrically modeling the realistic and exact human body shapes from intuitive and semantic linear anthropometric parameters such as height, weight. The modeling framework consists of the following three phases: construction of a training database of human body shapes, statistical analysis of human body shapes and human body shape modeling. In the training database construction phase, a consistent parameterization was carried out on 3D whole-body scan data of 80 males and 80 females with a wide variety of body shapes as the examples for this work. The surface-fitting process, which was improved relative to existing methods, was developed to guarantee the high-quality parameterization in this phase and to generate articulated body shape models in the modeling phase. To characterize the range of body shape variation, the training database was analyzed statistically. Additionally, a correlation between the body shapes and the sizes was learned in the statistical analysis phase to estimate new body shapes from intuitive and semantic anthropometric parameters. A new technique to generate an optimal body shape model that precisely satisfies user-input body dimensions was developed in the model generation phase. This technique enables the estimation of body shape variation, not only within the body shape space that was learned statistically, but also outside of the body shape space, while maintaining body shapes that stay in the human shape space. This work produced reasonable results having a high modeling accuracy satisfying user-specified anthropometric parameters, and high visual quality in expressing realistic body shapes. The resultant models were then segmented into 16 key regions of the human body, and had information on 15 key joints, and thus they could be a useful tool in various industries. The proposed method contributes to related areas by introducing an improved surface-fitting process and a non-linear optimization-based optimal body shape modeling technique. This thesis then introduces various applications using the statistical parametric human model developed in this work. As the first application, a posture modeling method basically required for human-centered product design in diverse industries is described. As the second application, a method for modeling an individual’s body shape is presented using Kinects and the statistical model so that the method can be used for personalized product design of advanced concept. And then, a method to predict the individual’s body type change, according to the adjustment of height and/or weight, is presented by learning the tendency of body shape variation that the statistical model has in the individual’s model. Lastly, texture mapping and morphing are demonstrated as a simple application used in the computer graphics field.