BTLD+:A BAYESIAN APPROACH TO TRACKING LEARNING DETECTION BY PARTS

The contribution proposed in this thesis focuses on this particular instance of the visual tracking problem, referred as Adaptive Ap- iv pearance Tracking. We proposed different approaches based on the Tracking Learning Detection (TLD) decomposition proposed in [55]. TLD decomposes visual tracking into three components, namely the tracker, the learner and detector. The tracker and the detector are two competitive processes for target localization based on comple- mentary sources of informations. The former searches for local fea- tures between consecutive frames in order to localize the target; the latter exploits an on-line appearance model to detect confident hy- pothesis over the entire image. The learner selects the final solution among the provided hypothesis. It updates the target appearance model, if necessary, reinitialize the tracker and bootstraps the detec- tor’s appearance model. In particular, we investigated different ap- proaches to enforce the TLD stability. First, we replaced the tracker component with a novel one based on mcmc particle filtering; after- wards, we proposed a robust appearance modeling component able to characterize deformable objects in static images; after all, we inte- grated a modeling component able to integrate local visual features learning into the whole approach, lying to a couple layered represen- tation of the target appearance.