Publication     Conference     Thesis

3D Human Motion Analysis

Motivation
  Computer systems are increasingly being used for sports training. Two kinds of computer-aided sports training systems
are commercially available: 3D motion-based systems and 2D video-based systems.

A 3D motion-based system uses multiple cameras to track the motion of reflective markers attached to the performer’s body. The markers’ 3D positions are recovered and used to compute the performer’s 3D motion, which can be analyzed by the coach or compared with a 3D reference motion of an expert. Such a system can provide accurate motion analysis. However, it is very expensive and difficult to use for the general users.

A 2D video-based system captures the performer’s motion using an off-the-shelf video camera and loads the video into a computer system. The system displays the performer’s video and a pre-recorded expert’s video side by side, and provides tools for the user to manually compare the performer’s motion with the expert’s motion. The system is affordable to general users. However, it cannot perform detailed motion analysis automatically.
   
Research Goal
  To overcome the shortcomings of existing systems, we propose a framework for affordable and intelligent sports training systems for general users that require only single stationary camera to record the user’s motion. In this framework, sports motion analysis can be formulated as a 3D-2D spatiotemporal motion registration problem. With this approach, the framework can be applied to analysis different types of sports motion using the appropriate 3D reference motion.
   
Sample Results
 
Taichi Motion  





input frames









frontal view
Blue: user's posture
Green: reference posture






side view
Blue: user's posture
Green: reference posture
   
Golf Swing Motion  




input frames









frontal view
Blue: user's posture
Green: reference posture







side view
Blue: user's posture
Green: reference posture
Video demo input-1, output-1
input-2, output-2
input-3, output-3
   
Researchers
>> Mr. Wang Ruixuan, Ph.D. student, Dept. of Computer Science, Natioanl University of Singapore
>> A/Prof. Leow Wee Kheng, Dept. of Computer Science, Natioanl University of Singapore
>> A/Prof. Leong Hon Wai, Dept. of Computer Science, Natioanl University of Singapore
>> Mr. Mark Lee, ResultSmith Corp.
>> Mr. Xing Dongfeng, R.A., Dept. of Computer Science, Natioanl University of Singapore
   
Patents
>> W. K. Leow, R. Wang, C.-S. M. Lee, D. Xing, and H. W. Leong. System and Method for 3D Human Motion Analysis from 2D Video. Provisional US patent filed on 9 Nov 2007.
   
Publications
>> R. Wang, W. K. Leow, and H. W. Leong. 3D-2D Spatiotemporal Registration for Sports Motion Analysis, Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 24-26 Jun 2008.
>> R. Wang and W. K. Leow. Human Posture Analysis under Partial Self-occlusion. In Proc. Int. Conf. on Image Analysis and Recognition, 2006.
>> R. Wang and W. K. Leow. Human Posture Sequence Estimation Using Two Un-calibrated Cameras. In Proc. British Machine Vision Conference, 2005.
>> R. Wang and W. K. Leow. Human Body Posture Refinement by Nonparametric Belief Propagation. In Proc. Int. Conf. on Image Processing, 2005.

3 July 2016