Video-based analysis of human motion

点击翻译按钮获取中文摘要

We spend a lot of time looking at people. Arguably, future computer systems will do so too, enabling new man-machine interfaces, perceptive consumer devices, smart surveillance systems, and myriad systems we have yet to envision. A key enabling technology for such purposes is video-based human pose tracking. The fidelity with which one needs to estimate 3D pose varies from task to task, but for most interesting applications current techniques are insufficient, especially when compared to the subtle movements with which humans communicate. The problems stem from the sheer number of joints in the body, which entails a high-dimensional estimation problem. Soft tissue and clothing obscure the skeleton, producing great variability in appearance, and further ambiguities are due to limited image resolution, the inability to distinguish which parts of an image correspond to which person, and because some body parts are occluded. To resolve ambiguities and reduce uncertainty caused by noisy (or missing) measurements, prior models play a key role in pose tracking, resolving an otherwise ambiguous estimation problem with a bias toward plausible human motions.