Full body modeling from video sequences

Synthetic modeling of human bodies and the simulation of mot ion is a longstanding problem in modeling and animation and much work has to be i nv stigated before a near-realistic performance can be achieved. At prese nt it takes an experienced graphic designer a long work-flow to build a complete and real istic model that closely resembles a specific person. Our goal is to automate the proce ss. This work pertains to the automation of synchronized multi-camera calibratio n and investigates in the reconstruction, modeling and tracking of human body models from video sequences. Such processes have many applications, as entertainment, s por s medicine/athletic training, and biometry. One topic of this work is the implementation of the fully auto matic geometric calibration software. The method needs a wand moving through object space. Dependent on the volume size and accuracy requirements different wands c an be used. Assuming a static multi station network every frame serves as snapshot for builing a well defined point cloud. We have developed a scene description language to achieve the most flexible way of data collection. Moreover, the data can be dis tributed over an internet network and different operating systems can be introduced. The idenfication of the points is established through standard point detection ope rators implemented through scripts, which are running locally on every workstation in t he intranet. These identified points together with the frame number are used to establi h relative orientations and to setup a network. These approximations are taken to set up a bundle adjustment network, allowing multiple command line switches to lead th e process. The approximations are first viewed in a OpenGL 3D viewer and can be analy sed for further runs or stopped, if an error occured. This first mile stone in t he geometric calibration procedure allows the user to inspect the correctness of the c ollection process. The unified bundle adjustment determines the network and output s all relevant geometric calibration protocols as text and graphical report. The image sequence character and the knowledge of additiona l spatial information about the “empty scene” and/or the human body approximation ll w to make useful combinations with useful results. The spatial and temporal ch racter of the video sequence is fully used. The foreground/background subtracti on algorithm is dependent on the base line of the multi camera setup. If there is a small b aseline between neighbouring sensors we use disparity maps to define the foregroun d and background. In an all around network with a small number of cameras where aut omatic identifica-