Visual tracking is broad area covering methods ranging from fast rectangular patch
matchers to complex articulated or non-rigid pose estimators. In recent years,
fueled by both progress in fast detection, learning and segmentation methods as
well as application-driven demand, many new tracking methods have emerged.
Published tracking methods differ in many aspects such as speed, complexity of the
model of the tracked entity, the (geometric) transformations assumed, mode of
operation (casual and non-causal), ability to adapt and learn, robustness to
occlusion and assumptions about the observer.
I will spend some time on various tracking formulation and show it is not easy to
define evaluate tracking performance and design informative benchmarks to assess
the multi aspects of tracker quality.
In the second part of the talk, I will present three trackers developed by me and my
collaborators that operate at very different points in the speed-robustness-flexibility
space. The TLD tracker (10 Hz, with redection and object appearance learning), the
Flock-of-Trackers (100 Hz, with robust pose estimation) and the Zero-Shift-Point
tracker (1000 Hz). I will focus on mechanisms for prediction and handling of tracking
errors. Such mechanisms contribute to tracker robustness.