Visual Tracking Using a novel feature pooling

It’s a challenging task to overcome problemsin visual tracking and create an appearance model thatmakes it possible to track a target robustly. Any videostream consist of a sequence of frames. At each framewhen the target is detected some positive and negativesamples are extracted around it. As a data independentbasis, wavelet is utilized to represent the target in differentscales. The wavelet transform of positive and negative arecomputed. A novel yet effective and efficient featurepooling is proposed to separate features that are a bitdifferent form each other, precisely speaking a smalldifference between two features results in a bigger onewhen they are pooled. Pooled features are fed to aBayesian classifier to update the classifier. Finally aparticle filter which its mean adoptively changes is used topredict the target location in the next frame and finds theareas that are more likely to be the target’s state, calledcandidates, and the samples that gain higher classifierscore are used to determine the target’s location.

Keywords:

Visual tracking , Compressed sensing , Particlefilter , Wavelet transform , Bayesian classifier