Loading...

3D Human Body Pose Estimation Using Multi-view Videos

Ramezanpour Namaghi, Sadegh | 2013

1838 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 44861 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Kasaei, Shohreh
  7. Abstract:
  8. Estimating 3D pose of the human body in videos has many potential applications in motion analysis, surveillance and human computer interfaces. In the last decade using more than one camera causes more accurate pose estimation. Most of the previous works use calibrated and fixed cameras but in action this is not a realistic assumption and using some arbitrary moving cameras is more likely. We introduce a novel method for this purpose. Our approach does not require specifying locations and directions of the cameras. The method uses a database of 3D poses for different human activities. This database includes body silhouettes from multiple viewpoints. To estimate the 3D body pose, silhouettes of the human body have been extracted from the videos, using background subtraction methods, and then we search them in our database and select some of the best samples. In the next step we recognize the action of the human using majority voting. After clustering the viewpoints of best database samples of each camera, the clusters centers are candidate viewpoints of the cameras. We check all combinations of the camera viewpoints and select the best of them that has the maximum consistency in all cameras, then estimate the pose from the multi-view images using the human action and cameras viewpoints. To use the temporal information of the video we keep the history of the actions and 3D poses of the previous frames and use them to reduce the search space. The method estimates the 3D pose of the human body using three views with a mean angular error less than three degrees between body joints
  9. Keywords:
  10. State Estimation ; Multiview Videos ; Articulated Body Model ; Silhouette ; Human Movment Analysis

 Digital Object List

 Bookmark

...see more