Motion capture based on RGBD data from multiple sensors for avatar animation

Abstract

With recent advances in technology and emergence of affordable RGB-D sensors for a wider range of users, markerless motion capture has become an active field of research both in computer vision and computer graphics. In this thesis, we designed a POC (Proof of Concept) for a new tool that enables us to perform motion capture by using a variable number of commodity RGB-D sensors of different brands and technical specifications on constraint-less layout environments. The main goal of this work is to provide a tool with motion capture capabilities by using a handful of RGB-D sensors, without imposing strong requirements in terms of lighting, background or extension of the motion capture area. Of course, the number of RGB-D sensors needed is inversely proportional to their resolution, and directly proportional to the size of the area to track to. Built on top of the OpenNI 2 library, we made this POC compatible with most of the nonhigh-end RGB-D sensors currently available in the market. Due to the lack of resources on a single computer, in order to support more than a couple of sensors working simultaneously, we need a setup composed of multiple computers. In order to keep data coherency and synchronization across sensors and computers, our tool makes use of a semi-automatic calibration method and a message-oriented network protocol. From color and depth data given by a sensor, we can also obtain a 3D pointcloud representation of the environment. By combining pointclouds from multiple sensors, we can collect a complete and animated 3D pointcloud that can be visualized from any viewpoint. Given a 3D avatar model and its corresponding attached skeleton, we can use an iterative optimization method (e.g. Simplex) to find a fit between each pointcloud frame and a skeleton configuration, resulting in 3D avatar animation when using such skeleton configurations as key frames.