Multi-resolution Shape Description And Clustering Of 3D Anthropometric Data For Population Fitting Design

Shape analysis and comparison is a fundamental part of many researches on anthropometry, biology, anthropology, archaeology, botany, and so on. In traditional methods based on linear distance, geometrical characteristics and internal spatial structure of a surface were not adequately considered. Three-dimensional (3D) anthropometry can provide rich information for ergonomic product design with better safety and health considerations. Shape comparison based on 3D anthropometric data is a complicated and challenging problem which has not been adequately studied and practiced yet. It is a promising field since it can lead to plenty of improvements in safe, healthy and comfortable civilian living and work, high military efficiency, etc.Multi-resolution description of 3D anthropometric data would facilitate the analysis. To reduce enormous computational load and memory requirement in 3D anthropometric data analysis, wavelet analysis is adopted in this paper to establish a multi-resolution mathematical description of 3D anthropometric data. This method provides flexible description of body surface shape at different resolution levels. Data analysis can then be performed with coarse resolutions, which preserve the major shape components but ignore micro shape components. A proper resolution can be selected according to the application purpose. The method mainly involves the transformation of arbitary bi-cubic surfaces to quasi-uniform B-spline surfaces and wavelet decomposition of quasi-uniform B-spline surfaces. To examine the feasibility of the multi-resolution description method in the description of the main shape of human body surfaces, upper head, whole head, and face samples have been analyzed. Approximation errors under different resolutions are presented. The application of the multi-resolution method in product design is also illustrated by examples.Given the wavelet decomposition of 3D data accomplished, k-means clustering is performed on the decomposed data to segment the population based on body surface shape. By using a novel block-based distance, each 3D head/face was transformed into a multi-dimensional vector. Not only the size information but also the geometric information of the 3D faces is consisted in this vector. Totally 378 face samples, 447 head samples and 432 upper head samples have been analyzed to illustrate the applicability and robustness of the method. Clustering validation was implemented by using two measures, i.e., size-weighted variances and Clustering Validity Indices (CVI). K-means clustering on different variables is compared, including head-length and head-breadth, top five principle components from Principle Component Analysis (PCA) on the proposed block distance-based vectors, and block distance-based vectors directly. The results show that the proposed block distance-based descriptor is superior to traditional sizing dimensions.The influence of three parameters (i.e., block number, resolution and alignment reference) on the clustering results was addressed. Results have shown that block number variation and resolution both have no obvious influence on the clustering results, while alignment reference has evident influence. Also comparative study was carried out on the clustering results by using the method proposed in this dissertation and by using the sizing tariff proposed by Chinese national military standard GJB 5477-2006. Results have shown their difference, which indicates the necessity to take shape information into consideration in the revision of the military standard in the future.The proposed sizing method incorporated with a multi-resolution representation of 3D anthropometric data and k-means clustering on the decomposed data provides a systematic method for proper grouping the samples into clusters according to their 3D shape.