Loading...

3D Human Pose Estimation

Ekhtiyari, Zahra | 2016

740 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 48018 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Kasaei, Shohreh
  7. Abstract:
  8. During past decade, human pose estimation has become a very important topic in computer vision with many applications; such as robotic, human computer interaction, and analyzing sport videos. Pose estimation is the process of estimating the configuration of the body joints from one or more images. Pose estimation in 2D space has been performed by processing of the images of a single camera. Recently, the multi-view methods have been emerged to estimate 3D human pose.The purpose of this thesis is 3D human pose estimation from multi-view images, captured in football matches. Pictorial structure is an efficient method in 2D human pose estimation. The 3D extension of this method is studied in this thesis. In the proposed method, at first the 2D human pose is estimated, , then inaccurate initial 3D pose of human is computed using 3D reconstruction. The 3D search space for each joint is limited by sampling from combination of several Gaussian distributions centered at that joint. One of the challenges in 3D extension of 2D pictoral structure is the large search space due to the degrees of freedom of human pose estimation in 3D space, which is handled in this method to reduce the run time.The skeletal human model is considered as a tree structure whose nodes are human joints. The main idea of the proposed method is that the distance between joints in 3D space should be proportional to human body configuration. This key condition is only feasable in 3d space whereas due to camera projection in 2D space, it is violated. For solving this problem, a conditional random field with unary and pairwise terms is used. Inference is computed using max product algorithm to obtaine optimal human body pose.For comparing the proposed method with other methods a percentage of correctly detected pose is used. In comparison with other researches, the proposed method increases the PCP about 20 %. Moreover, compared with other methods, the runtime of this method is significantly reduced
  9. Keywords:
  10. Multiview Cameras ; Human Pose Estimation ; State Estimation ; Three Dimentional State Estimation ; Pictorial Structure ; Max Product

 Digital Object List

 Bookmark

No TOC