Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (20): 210-220.DOI: 10.3778/j.issn.1002-8331.2102-0098

Previous Articles     Next Articles

Object Tracking with Anchor-Free and Online Updating

ZHANG Rui, SONG Jingzhou, LI Sihao   

  1. School of Automation, Beijing University of Posts and Telecommunications, Beijing 100876, China
  • Online:2021-10-15 Published:2021-10-21



  1. 北京邮电大学 自动化学院,北京 100876


SiamRPN is an anchor-based tracking algorithm, which is not robust to scale change, severe deformation, and rotation. This paper proposes a target tracking algorithm with anchor-free and online updating. Firstly, a multi-layer fusion feature extraction network is proposed, which can make full use of the structure and semantic information of the image. Secondly, an anchor-free mechanism is adopted, which enables the network to directly predict the distance from the sampling points in the target area to the boundary of the target area. The anchor-free mechanism can avoid the shortcomings of the anchor-based mechanism. Finally, an online updating module is added to the backbone network. The online updating module uses the latest tracking results for online training, so that the algorithm can better predict targets that do not appear in the training set, furthermore it can make the algorithm adapt to the change of the targets during tracking. Compared with SiamRPN algorithm, the success rate and accuracy rate of this algorithm on OTB100 data set are increased by 0.062 and 0.065. It shows better robustness to scale change, severe deformation and rotation.

Key words: multi-layer fusion, anchor-free mechanism, online updating, Siamese network



关键词: 多层特征融合, 无锚点机制, 在线更新, 孪生神经网络