IS = { zkontrolovano 15 Jan 2013 },
  UPDATE  = { 2012-10-01 },
author =      {Voj{\'\i}{\v r}, Tom{\'a}{\v s}},
supervisor =  {Matas, Ji{\v r}{\'\i}},
title =       {Long-term Visual Object Tracking with Online Learning -- {PhD} Thesis Proposal},
institution = {Center for Machine Perception, K13133 FEE
               Czech Technical University},
address =     {Prague, Czech Republic},
year =        {2012},
month =       {August},
day =         {},
type =        {Research Report},
number =      {CTU--CMP--2012--18},
issn =        {1213-2365},
pages =       {32},
figures =     {13},
authorship =  {100},
psurl       = {[Vojir-phdproposal-2012.pdf]},
annote =      {In this report, we outline recent work in the visual
    object tracking. Especially, we focuse on the
    Tracking-Learning-Detection (TLD) algorithm which combines
    frame-to-frame object tracking, object detection in the whole
    image and online learning of the object appearance.  The report
    presents a contributions to the design of the TLD frame-to-frame
    tracker. The new tracker, called Flock of Trackers (FoT),
    estimates the pose of the tracked object by robustly combining
    displacement estimates from a subset of local trackers that cover
    the object. The first contribution, called the Cell FoT, allows
    local trackers to naturally drift to good points to track. As a
    second contribution, we introduce two new predictors of local
    tracker failure -- the neighbourhood consistency predictor and the
    Markov predictor -- and show that the new predictors combined with
    the normal- ized cross-correlation predictor are more powerful and
    almost two times faster than the baseline TLD tracker predictor
    based on normalized cross-correlation and the forward- backward
    procedure. Finally, two methods for combining the individual
    predictors are proposed. The first that combines the predictors by
    the AND operator and the second that learns the likelihood of
    being a inlier. The new FoT was compared with the baseline tracker
    and outperforms it in the processing speed, robustness and overall
    tracking performance on almost all test sequences. A comparison
    with the state-of-the-art tracking algorithms shows that the new
    FoT achieves the best average performance. At the end, a further
    research directions are discussed, which points out some of the
    shortcomings of the TLD framework and how can they be overcome. },
keywords =    {tracking,long-term,object detection},