We have shown previously that it is possible to accurately reconstruct periodic motions in 3D from a single camera view, using periodicity as a physical constraint from which to perform geometric inference. In this paper we explore the suitability of the reconstruction techniques for real human motion. We examine the degree of periodicity of human gait empirically, and develop algorithmic tools to address some of the challenges arising from this type of motion, including reconstructing motions that deviate from pure periodicity, properly handling the trajectories of multiple points on an articulated body, and proposing a distance function for measuring the difference between two reconstructions. Importantly, we illustrate the usefulness of these techniques by applying them to the tasks of view-invariant activity classification, clinical gait analysis and person identification.
Bibliographical noteFunding Information:
The authors wish to thank Professors David Nuckley and James Carey from the Department of Physical Therapy at the University of Minnesota for providing access to the stroke patient gait data used in this paper. In addition, this material is based upon work supported in part by the National Science Foundation through Grants #IIP-0443945 , #CNS-0821474 , #IIP-0934327 , #CNS-1039741 , and #SMA-1028076 . The data for subject identification experiments were obtained from CMU Graphics Lab Motion Capture Database ( http://mocap.cs.cmu.edu/ ), funded by NSF EIA-0196217 , and the Georgia Tech ‘Human Identification at a Distance’ Database, funded by the DARPA HID Program.
- 3D reconstruction
- Activity classification
- Gait analysis
- Human motion