Tracking background

The consortium's experience in tracking systems is to a large extend consolidated in the book  Audio-Visual Person Tracking: A Practical Approach (World Scientific Publication Co., Imperial College Press, London, UK, Dec. 2011).

SMART extentions

The SMART project builds upon the algorithmic background of particle filtering. The SMART face and person tracking systems differ from the existing versions in two aspects:

• The measurement cues yielding the likelihood of measurement given the state of the tracker are enhanced.
• Different measurement fusion strategies are explored.

Measurement cues

• Motion: Combination of short-term (frame-by-frame difference) and medium-term (difference from estimated background).
• Colour: Dependent colour components in histograms that attenuate the importance of imminent backgound, either around the target or in the background estimation behind it.
• Detection: The likelihood of face but also body detection is calculated based on how many stages in a cascaded classifier verify the object presence and with what margin has this verification happened.

Fusion strategies

• Early fusion: Likelihood combination.
• Mid fusion: Fusion within the tracker, utilising partitioned sampling.
• Late fusion: Combining outputs of trackers utilising each ones track confidence.

Video demonstrations

Jan 2012: Face tracker employing colour, motion and face detection measurements, and their low-level fusion by likelihood multiplication. The face detection measurements are used for initialisation.