Seminar: Real-time 3D Registration of Multiple Depth Cameras

Sahand Seifi
M.Sc. Thesis Proposal
Supervisor: Dr. Oscar Meruvia-Pastor

Real-time 3D Registration of Multiple Depth Cameras

Department of Computer Science
Friday, October 4, 2013, 11:30 a.m., Room EN 2022


Depth-sensing cameras have been widely commercialized and available for industry, research and entertainment in the past few years. By providing the depth of objects in a scene, they allow us to use computers to build and process 3-Dimensional models of the world in real-time, rather than relying on 2D images from conventional cameras. This has led to breakthroughs in computer vision, human-computer interaction and entertainment industry, and has been applied to many areas such as: vehicle and robot navigation, creation of CGI animations and video game industry. One of the challenges in using such technology in real-time applications is when the depth camera moves and future inputs should go through a transformation, based on the camera rotation and translate, before they conform and are applied to the already built model. The required transformation can be estimated by the computer by comparing the input with the model. This process of merging imagery inputs of different viewpoints or times into a single model is called registration. The purpose of this research is to merge the input streams of multiple moving depth cameras into one model in real-time. Registration algorithms are time consuming, so we aim to implement and improve these algorithms to the point that they can be used in real-time with constraints specific to our application. We will also look into effective down-sampling algorithms and use them to decrease the size of input to registration algorithms.