THESIS
2017
Abstract
Human shape information is useful for many applications, ranging from entertainment, e-commerce,
big data research to biomedical research. More recently due to strong growth of
online shopping for fashion industry, there are even needs for collecting body shape
information from customers for customization and size recommendation purposes. Collection
of body dimensions is not a trivial task. People may not possess knowledge and skills to
obtain measurements correctly by themselves. Traditionally the task is carried out by a tailor
on site. The introduction of 3D scanner allows automatic collection of detailed body
geometry information. However the subject is required to wear tight fitting clothes during the
scanning process. Also scanning is carried out at specific location. Some...[
Read more ]
Human shape information is useful for many applications, ranging from entertainment, e-commerce,
big data research to biomedical research. More recently due to strong growth of
online shopping for fashion industry, there are even needs for collecting body shape
information from customers for customization and size recommendation purposes. Collection
of body dimensions is not a trivial task. People may not possess knowledge and skills to
obtain measurements correctly by themselves. Traditionally the task is carried out by a tailor
on site. The introduction of 3D scanner allows automatic collection of detailed body
geometry information. However the subject is required to wear tight fitting clothes during the
scanning process. Also scanning is carried out at specific location. Some people may find it
intrusive and inconvenient.
In this research, the possibility of using monocular sequence for human body shape
information is demonstrated. A framework of highly automated non-intrusive method for
obtaining pose and shape information of a dressed person from monocular gaiting sequence is
proposed. A synthesis and test approach using a statistical human model is adopted. The pose
of each frame in gaiting sequence is first initialized based on shape context matching.
Temporal constraint is introduced to avoid error due to shape ambiguity. A pose and shape
refinement is performed to bring initial estimation closer to optimal solution. This addition
step reduces search space and thus computation time for later pose and shape optimization.
Pose and shape optimization is based on annealed particle filtering. An error model for single
view estimation is proposed to obtain body dimensions of dressed person. Experiments were
carried out on the prototype system and analysis of the result is presented.
Post a Comment