CAREER: Acquiring Computer Graphics Models of Human Shape and Motion

点击翻译按钮获取中文摘要

AbstractAcquiring Computer Graphics Models of Human Shape and MotionJames Davis, UC Santa CruzThis research is developing algorithms to measure human body shape and motion to 1mm and 1deg accuracy without placing "fiducials" (i.e., markers) on the person's body. The major advances in this research over previous efforts to do accurate human motion tracking are replacing traditional cameras with high accuracy 3D shape measurement devices and utilizing a carefully constructed body model. Improved technology for the reliable and accurate measurement of human movement will ultimately enable new applications in ergonomics, smart spaces, fashion, surveillance, surgery, security, health, user interfaces, and art. The project uses two specific application areas for evaluation of the scientific goals: improving knowledge about the causes and remedies of athletic injuries through direct measurement of motion and archiving dance performances as an educational resource.This research transforms human shape and motion tracking from a problem of optimization to a problem of human shape priors and real-time, 3D shape measurement. This requires 3D shape sensors which are accurate, work with moving subjects, and can support multiple simultaneous viewpoints. Existing structured light triangulation and time-of-flight sensors do not support these requirements; however, they do have complementary qualities and deficiencies. The investigators study two novel designs that combine these modalities to achieve the improved accuracy. The method requires a prior model consisting of a very detailed and accurate model of human surface shape as a function of pose and identity. This model is constructed initially from full body laser scans of a variety of people with different body types in a variety of different poses. When the highly accurate real-time 3D data measurement system has been fully developed, a much larger training data set will be captured. Improvements in data measurement and prior models will provide a better conditioned optimization space with fewer local minima, allowing for more robust and accurate estimation.