Human Body Model Acquisition and Tracking using RGB-D camera

Acquisition and tracking of human body is a key research problem in the fields of computer vision, computer graphics, and biomechanics. We propose a system that can used to create personalized human body model in normal room environment using just a single RGB-D camera. We want our system to be robust against illumination variations and complex background. Our system first estimates user specific full body model by registering a template body mesh to a point cloud estimated by the depth camera. We track important landmarks of human body visible in the video using large displacement optical flow techniques. The corresponding 3D landmarks in the template mesh are annotated and projected to the image plane using camera projection matrix. Using an optimization technique we correct the locations of the projected markers by minimizing their deviation from the tracked 2D landmarks. Based on these corrections we update registered mesh which can be used as a personalized avatar of the user. The system will be designed to keep the requirement of manual processing, such as camera calibration and landmark selection, to a minimum.

Faculty Supervisor:

Dinesh Pai

Student:

Partner:

Université Grenoble Alpes

Discipline:

Computer science

Sector:

Education

University:

The University of British Columbia

Program: