Robust Patch-based Object Tracking Via Superpixel Learning

With the rapid deveopment of information technology, the demand forvideo image processing increases. Object tracking has gained a wide-ranging applications in surveillance, augmented reality, human-computerinteraction, triggering a futher exploration and extensive research in thefield of object tracking. Many scholars have proposed numerous theoriesand methods for object tracking. But as to the complexity and variability inreal-world tracking scenarios, object tracking still faces enormouschallenges. The purpose of our research is to design an effective andreliable object tracking system, being robust of severe object deformationand different changes in illumination.In this paper, different from the computation of traditional histograms,a novel histogram is proposed which takes the spatial position of pixelsinto consideration when computing. Based on the improved histogram, wefurther introduce a calculation method of illumination invariant feature foreach pixel. Then we present the concept and computing of superpixel, andgenerate foreground/background confidencemap by superpixel learning.For object tracking under severe deformation, we propose a patch-based object tracking framework. Patches are selected via superpixelsegmentation, then we build a star topology model to describe relationshipbetween the target region and the patches. Under the overall framework ofparticle filter tracking, we use adaptive Basin Hopping Monte Carlomethod for sampling. And a co-training framework is proposed wheresuperpixel learning is used to update the patch appearance and vice versa.For the case of illumination changes, we firstly transform the pixel valueinto illumination invariant features, and then utilize the patch-based target tracking system to do the follow-ups. Experimental results show that ourproposed patch-based tracking system can achieve good trackingperformance under severe deformation and illumination changes.