中文版 | English
CSpace  > 智能工业设计工程中心
Collaborating visual tracker based on particle filter and correlation filter
Li, Weiguang1,3; Wei, Wang2; Qiang, Han3; Shi, Mingquan1
2019-06-25
摘要Correlation filter (CF)-based tracking algorithms is most popular in recent years due to its high accuracy and impressive speed. However, it has some intrinsically drawbacks such as margin suppression, sensitive for disturbance, and partial occlusions. Contrasted with CF drawbacks, the advantages of particle filter (PF) tracking algorithm include robustness, motion prediction, and wide detection range. Therefore, it can amend some CF tracker drawbacks. On the other hand, the HOG feature is widely used in CF tracker because it can detect the target precision position. However, this kind of feature is rotation-variation, which is invalid for rotation transformation target. On the contrary, the tracker precision merely based on colour feature is rough, but colour feature is rotation invariation and is effective for rotating target; therefore, these two features are complementary. In this paper, we integrate both trackers (CF and PF) to learn the HOG and colour feature, respectively, experiments demonstrate this tracking algorithm is more robust, and the tracking precision is more accurate. This algorithm is integrated with some classic CF trackers (KCF, SAMF, and MOSSE) framework and benchmark them against their baseline. On the OTB2015 benchmark datasets, experiment result demonstrates OPE performance grades have improved from about 1% to 12%; SRE Performance grades have improved from about 1.3% to 5.8%.
关键词Bayesian estimation computer vision correlation filter particle filter
DOI10.1002/cpe.4665
发表期刊CONCURRENCY AND COMPUTATION-PRACTICE & EXPERIENCE
ISSN1532-0626
卷号31期号:12页码:13
收录类别SCI
WOS记录号WOS:000468975000014
语种英语