[Univ of Cambridge]

Edge-based Tracking

[Dept of Engineering]

The approach is based upon a CAD model of the structure that is to be tracked.

The CAD model is rendered from a predicted viewpoint (this may be where the object was found in the previous frame, or a motion estimate may be used to give a more accurate prediction).

This identifies the edges of the model which are expected to be visible. Sample points are then assigned at uniform intervals along these visible edges
A search is then conducted from each sample point for an intensity discontinuity in the video image

A change in the pose estimate is then computed which minimises these errors. This is achieved using iterative re-weighted least squares which allows the use of an M-estimator which is robust to outliers. The following videos show the basic tracker in action.


These videos were made when the tracker was first implemented on an SGI with a 225MHz R10K CPU
  • An mpeg movie of the tracking system (~1MB)
  • Movie showing basic robustness to occlusion (~1MB)


    Real-time tracking of complex structures with on-line camera calibration (BMVC 1999)
    T.W. Drummond
    Last modified: Tue Feb 20 15:24:34 GMT 2007