High performance rigid body tracking

Many potential applications of computer vision require a 3D tracking system which can cope with very unpredictable, rapid motions of the camera while operating at frame rate. This cannot be achieved using previously available techniques. Robustness in tracking is achieved by combining different systems in a non trivial way. This allows trackers to be designed in such a way that they are not general purpose trackers, but instead excel at some other chosen property. The problem of robustness is addressed by the development of a point based tracking system which uses high speed techniques for point detection and matching (allowing it to operate on full size frames) combined with a robust optimizer. This, coupled with a system which estimates the quality of a match, allows the system to track very rapid, unpredictable motions with considerable amounts of noise. The full-frame extraction of feature points is vital to the robustness of the system. Machine learning is used to derive a feature detector which is significantly faster than previous methods. The repeatability of extracted features has been verified by comparison to other detectors. Despite being principally constructed for speed, the detector significantly outperforms existing feature detectors. The problem of tracking curved surfaces is approached by developing a scheme for rapidly rendering the apparent contour from a predicted pose (a step required for tracking). This is done efficiently with a method for quickly tracing out the apparent contours, combined with family of high speed polygon intersection algorithms and a set of rules which determine visibility from the contour intersections. Finally, a careful analysis of the properties of the trackers is performed. This is used to design a non-trivial filter for combining the measurements, and it shown that the design choices collectively lead to a very high performance tracker.