PROJECT TITLE :

Parallel Attentive Correlation Tracking

ABSTRACT:

There is evidence to suggest that visual attention and selection in humans may be processed in simultaneously, based on psychological and cognitive results. This research provides a unique correlation filter (CF)-based tracking strategy, which corresponds to processing local and semi-local background domains, respectively, by reflecting various elements of these attentions. An efficient Boolean map format, inspired by Gestalt principles of figure-ground segregation, characterises an image by a series of Boolean maps via randomly thresholding its colour channels, providing a location response map as a weighted sum of all Boolean maps. In order to better tracking of non-rectangular objects, Boolean maps capture the target's and its scene's topological structures at multiple granularities. Another approach is to use new distractor-resistant metric regularizations in CF in the semi-local domains, which operate as a force to drive distractors towards the negative. CF's border effects can be successfully mitigated as a result. A Bayesian framework is used to combine the local and semi-local domain models, and the tracked location is selected by maximising the likelihood function of this framework. On a single CPU, extensive tests using the tracking benchmarks OTB50/OTB100/VOT2016/VOT2017 show that the suggested approach outperforms other cutting-edge trackers with a speed of 45 fps.