Research Article
Accurate Scale-Variable Tracking
@INPROCEEDINGS{10.1007/978-3-319-73564-1_5, author={Xinyou Li and Wenjing Kang and Gongliang Liu}, title={Accurate Scale-Variable Tracking}, proceedings={Machine Learning and Intelligent Communications. Second International Conference, MLICOM 2017, Weihai, China, August 5-6, 2017, Proceedings, Part I}, proceedings_a={MLICOM}, year={2018}, month={2}, keywords={Correlation tracking Scale estimation CNN features}, doi={10.1007/978-3-319-73564-1_5} }- Xinyou Li
Wenjing Kang
Gongliang Liu
Year: 2018
Accurate Scale-Variable Tracking
MLICOM
Springer
DOI: 10.1007/978-3-319-73564-1_5
Abstract
In recent years, several correlation tracking algorithms have been proposed exploiting hierarchical features from deep convolutional neural networks. However, most of these methods focus on utilizing the CNN features for target location and neglect the changes of target scale, which may import error to the model and lead to drifting. In this paper, we propose a novel scale-variable tracking algorithm based on hierarchical CNN features, which learns correlation filters to locate the target and constructs a target pyramid for scale estimation. To evaluate the tracking algorithm, extensive experiments are conducted on a benchmark with 100 video sequences, which demonstrate features exploited from different CNN layers are well fit to estimate the object scale. The evaluation results show that our tracker outperforms the state-of-the-art methods by a huge margin (+14.6% mean OS rate and +14.3% mean DP rate).