Defense Vasileios Belagiannis: Human Pose Estimation in Complex Environments
 |
- Speaker: Vasileios Belagiannis
- Date: Wednesday, September 2, 2015
- Time: 1:00 PM
- Location: FMI-Building, Room 03.13.010
|
Abstract
Estimating human body poses from images is a demanding task that has attracted great interest from the computer vision community. Determining automatically the body pose promotes many applications such as human tracking, motion capture, activity recognition, surveillance and surgical workflow analysis. This work addresses the problem of human pose estimation from different perspectives. At first, we tackle the problem of single human pose estimation from a single view. Then, we move to multi-view camera systems, where we work on both single and multiple human pose estimation. In this thesis, we propose novel discriminative and generative methods to address all these problems of human pose estimation.
The primary contributions of this work are threefold. At the beginning, we propose two discriminative methods for 2D human pose estimation from a single view. The first contribution exploits Random Forests (RF), while the second builds on Deep Learning. In both cases, we formulate the problem of body pose estimation as a regression task, where the body pose is defined by a set of body joints. To build a regressor that predicts the 2D body poses, we learn a model either using a Regression Forest or Convolutional Neural Network (
ConvNet?). Our third contribution is a generative model for multi-view human pose estimation. In particular, we address the problem of 3D pose estimation of multiple humans from multiple views and propose a 3D pictorial structures (3DPS) model. The model is generic and applicable to both single and multiple human pose estimation. Finally, we apply the 3D pictorial structures (3DPS) model on estimating the body pose of multiple individuals from multiple cameras in the operating room (OR). Therefore, this work considers several aspects of the human pose estimation problem, starting from single view scenarios and completing with human pose estimation multiple views in complex environments.
The talk will be in English and is open to the public. Everybody is cordially invited to attend.
