dc.description.abstract | Visual tracking of arbitrary objects is an active research topic in computer vision, with applications across multiple disciplines including video surveillance, activity analysis, robot vision, and human computer interface. Despite great progress having been made in object tracking in recent years, it still remains a challenge to design trackers that can deal with difficult tracking scenarios, such as camera motion, object motion change, occlusion, illumination changes, and object deformation. A promising way of tackling these types of problems is to use a part-based method; one which models and tracks small regions of the object and estimates the location of the object based on the tracked part's positions. These approaches typically model parts of objects with histograms of various hand-crafted features extracted from the region in which the part is located. However, it is unclear how such relatively homogeneous regions should be represented to form an effective part-based tracker. In this thesis we present a part-based tracker that includes a model for object parts that is designed to empirically characterise the underlying colour distribution of an image region, representing it by pairs of randomly selected colour features and counts of how many pixels are similar to each feature. This novel feature representation is used to find probable locations for the part in future frames via a Bhattacharyya Distance-based metric, which is modified to prefer higher quality matches. Sets of candidate patch locations are generated by randomly generating non-shearing affine transformations of the part's previous locations and locally optimising the most likely sets of parts to allow for small intra-frame object deformations. We also present a study of model initialisation in online, model-free tracking and evaluate several techniques for selecting the regions of an image, given a target bounding box most likely to contain an object. The strengths and limitations of the combined tracker are evaluated on the VOT2016 and VOT2018 datasets using their evaluation protocol, which also allows an extensive evaluation of parameter robustness. The presented tracker is ranked first among part-based trackers on the VOT2018 dataset and is particularly robust to changes in object and camera motion, as well as object size changes. | en_GB |