Structure and Motion Analysis

The term "tracking" has been used in various contexts. In Computer Vision, the analysis of video sequences can lead to the "tracking" of 2D image objects over time. Examples are the tracking of people or cars using stationary surveillance cameras. Many approaches have been proposed, including optical flow, particle filters, and mean-shift. The other major direction in tracking research is towards metric recovery of "structure and motion", i.e. reconstruction of the (static) 3D scene, and recovery of the full six degrees of freedom (6 DoF) of camera pose. Especially the problem of robust real-time head (camera) pose is very significant for Augmented Reality and Robotics applications. Various solutions have been proposed for vision-based pose tracking from artificial landmarks. There are also impressive recent results for online structure and motion analysis. However, in completely unknown environments, purely vision-based algorithms tend to drift, so that re-initialisation is required. Our own work is based on the fusion of visual point-features (corners) and inertial sensors. In both cases, visual features have to be extracted a priori from a known 3D scene model, and the update (maintenance) of scene models would require online structure and motion analysis.

Real-time Structure and Motion Estimation:

In this research area, we are focused on structure and motion estimation in order to navigate in unknown and unprepared environment. This task can be described as follows:
1) as the system navigates, it tracks some image points with calibrated cameras,
2) then the system is able to estimate a coarse motion of the camera w.r.t. the scene,
3) finally, the structure (i.e. the geometry) of the scene can be estimated thanks to the already known motion of the cameras.
This process is repeated until the convergence of the system, i.e. the structure of the scene and the motion of the camera are accurately recovered.

Contact: Axel Pinz, email: axel.pinz@tugraz.at