Human Body ASM Modeling And Its Application In Virtual Fitting

With the rise and development of Internet technology, apparel e-commerce has gradually developed into a mature market and become an important channel for clothing product sales and promotion. Currently, the defect of apparel e-commerce is that people can not judge fitting effect before their buying. Therefore, there is a broad market space for network clothing customization system which can meet the need of the virtual fitting. User’s size numbers are important in network clothing customization system. It is a hard technical problem for how to figure out and automatically locate the body feature points, access remote user’s size numbers and show the virtual fitting effect.To solve these problems, we study Active Shape Model(ASM) modeling and virtual fitting technology based on the 2013 Natural Science Foundation project of Science and Technology Department in Shaanxi Province, "Human body, clothing PDM(Point Distribution Model) modeling and its applications for virtual fitting "(2013JM8034). The main work includes:(1) Human body modeling and feature point positioning method based on Active Shape Model(ASM) is proposed. We establish the training sample set of the human body data which obtained via 3D body scanner. The feature points of human body image are extracted, and the training sample set is aligned and analyzed using principal component analysis. Human body Point Distribution Model(PDM) is built, and shape parameters is obtained, which would control the overall body shape deformation parameters. The maximum and minimum range of the human body feature regions are received, which provided a scientific basis for follow-up clothing regional deformation.(2) ASM local texture model is established, it is corresponding with every feature point in human training sample images. The gray scale distribution of each feature point on the body image is gained by statistical analysis, and a search algorithm is used to achieve the body feature point automatic positioning, access to the user’s body shape parameters.(3) Traditional ASM modeling and feature point location algorithm is effectively improved, and the effect of the algorithm before and after being improved are conducted by comparative analysis. Since traditional Procrustes normalization needs many iterations which will spend a lot of time, here training samples alignment is set by marking anchor point and using average body shape as the initialization rules model. Meanwhile, faced with the defect that traditional ASM method equating each candidate point of feature points in the neighborhood, which led to the two feature points indistinguishable for gray model’s similarity, locally weighted statistical model for candidate points is introduced, and the "coarse-to-fine" multi-resolution search method is used. Experimental results show that the feature point positioning method for Human body based on our improved ASM reduces the number of iterations, shortens the running time, and improves the positioning accuracy.(4) According to the obtained user’s body shape parameters, a prototype garment is used as basis template, on the foundation of each characteristic region is divided automatically, we utilize linear interpolation algorithm for clothing regional deformation, and ultimately achieve the effect of virtual fitting compliance with the user’s own body feature.