Research On Visual Tracking And Scene Illumination Based On Per-exemplar Learning

The idea of exemplar representation is to use the exemplars which are similar to the target in the dataset to represent the target, which is consistent with the mechanism of human visual expression, rather than dividing the dataset into distinct categories. Moreover, exemplar representation does not need large amount of data, thus it is simpler than the issue of category. Such technology has been successfully applied to object detection, scene parsing and other computer vision tasks. In this thesis, we extend exemplar representation to visual tracking by hard negatives mining. Further, the color of the object may change when the target undergoes dramatic illumination variation (leading to tracking drift). We attempt to employ exemplar representation to learn the illumination model for each surface and estimate the illumination. This would be benefical to solve the illumination variation of object tracking.This thesis includes the following three aspects:(1) We study the basic theory of the exemplar representation and introduce the visual similarity measure and learning process of the exemplar representation. Moreover, we analyse the applications and development of the theory in computer vision community.(2) For the requirement of large datasets and the difficulty of online update of tracking by detection, we extend exemplar representation to the field of tracking. We use a single positive exemplar and certain hard negatives to train an exemplar model. Meanwhile, we retain a specific number of the historical templates to correct the current template to refine tracking results.(3) For the illumination variation of object tracking, we attempt to use exemplar representation to learn the illumination model for each surface to estimate the illumination. Furthermore, we use the diagonal model to transform the estimated illumination into standard illumination to reduce the influence of illumination.