3D Head Rigid Motion Estimation

3D Head Model

Aim

Estimate the 3D rigid motion of a moving head from a monocular sequence.

Description of the problem

Let's assume a person that faces a camera and occasionally moves his head. The real motion of the head can be analysed to rigid and non-rigid motion. The rigid motion can be described by 6 parameters: a 3 dimensional translation vector T and a 3x3 rotation matrix R with 3 degrees of freedom:
X'=RX+T
The problem that was tackled was to estimate the parameters of the 3D motion for every pair of consecutive frames.

Applications

A automated system that can estimate the 3D motion parameters from a monocular sequence of a moving head can be used for telecommunication applications (e.g. video-conference using MPEG4 compression) or entertainment purposes (controlling virtual 3D characters)

Approach

The 3D rigid motion parameters are estimated using a two second approach. First, a set of image features are tracked and for each feature and each pair of consecutive frames, a 2D translation vector is estimated. Then, for each frame, the 3D motion parameters are calculated, using all the feature translation vectors. The features are linked to 3D head model which is assumed adapted at the first frame of the video sequence. It is also assumed that only rigid motion is observed.
The features are tracked using the Kanade-Lucas-Tomasi (KLT) algorithm. Features are assessed according to their trackability, dissimilarity and 3D orientation.




An initial estimate of the 3D motion parameters is calculated by an enhanced version of the 8-point algorithm. A final estimate is provided by the iterative algorithm Levenberg-Marquardt that minimises the distances between the tracked features and the projections of the their associated positions of the 3D model.
The motion parameters are applied to the 3D model and its projection is drawn onto the image to validate the result. Finally, the facial features are re-positioned according to the projection of the 3D model.




A summary of the method is given by the following diagram

Method Review diagram

Publications

N. Sarris, D. Makris, M. G. Strintzis, "Three Dimensional Model Based Rigid Tracking of a Human Head", International Workshop on Intelligent Communication Technologies and Applications with Emphasis on Mobile Communications, May, Neuchatel, Switzerland, (1999) abstract download

About this work

The above summarises Dimitrios Makris' final year thesis for the Greek Diploma. It took place at Aristotle University of Thessaloniki, Greece, during the academic year 1998-1999.