This thesis investigated single-target tracking of arbitrary objects. Tracking is a difficult problem due to a variety of challenges such as significant deformations of the target, occlusions, illumination variations, background clutter and camouflage. To achieve robust tracking performance under these severe conditions, this thesis proposed firstly a novel RGB single-target tracker which models the target with multi-layered features and contextual information. The proposed algorithm was tested on two different tracking benchmarks, i.e., VTB and VOT, where it demonstrated significantly more robust performance than other state-of-the-art RGB trackers. Proposed secondly was an extension of the designed RGB tracker to handle RGB-D images using both temporal and spatial constraints to exploit depth information more robustly. For evaluation, the thesis introduced a new RGB-D benchmark dataset with per-frame annotated attributes and extensive bias analysis, on which the proposed tracker achieved the best results. Proposed thirdly was a new tracking approach to handle camouflage problems in highly cluttered scenes exploiting global dynamic constraints from the context. To evaluate the tracker, a benchmark dataset was augmented with a new set of clutter sub-attributes. Using this dataset, it was demonstrated that the proposed method outperforms other state-of-the-art single target trackers on highly cluttered scenes.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:687469 |
| Date | January 2016 |
| Creators | Xiao, Jingjing |
| Publisher | University of Birmingham |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://etheses.bham.ac.uk//id/eprint/6688/ |
Page generated in 0.0023 seconds