Publication Type

Journal Article

Version

publishedVersion

Publication Date

4-2014

Abstract

The appearance model has been shown to be essential for robust visual tracking since it is the basic criterion to locating targets in video sequences. Though existing tracking-by-detection algorithms have shown to be greatly promising, they still suffer from the drift problem, which is caused by updating appearance models. In this paper, we propose a new appearance model composed of ranking middle-level patches to capture more object distinctiveness than traditional tracking-by-detection models. Targets and backgrounds are represented by both low-level bottom-up features and high-level top-down patches, which can compensate each other. Bottom-up features are defined at the pixel level, and each feature gets its discrimination score through selective feature attention mechanism. In top-down feature extraction, rectangular patches are ranked according to their bottom-up discrimination scores, by which all of them are clustered into irregular patches, named ranking middle-level patches. In addition, at the stage of classifier training, the online random forests algorithm is specially refined to reduce drifting problems. Experiments on challenging public datasets and our test videos demonstrate that our approach can effectively prevent the tracker drifting problem and obtain competitive performance in visual tracking.

Keywords

Middle-level Patches, Selective Feature Attention, Random Forests, Tracking-by-detection

Discipline

Computer and Systems Architecture | Computer Engineering

Research Areas

Data Science and Engineering

Publication

International Journal of Advanced Robotic Systems

Volume

11

First Page

1

Last Page

9

ISSN

1729-8806

Identifier

10.5772/58430

Publisher

SAGE Publications (UK and US): Open Access Titles / SAGE Publishing

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.

Additional URL

https://doi.org/10.5772/58430

Share

COinS