东北大学学报(自然科学版) ›› 2023, Vol. 44 ›› Issue (9): 1227-1233.DOI: 10.12068/j.issn.1005-3026.2023.09.002

• 信息与控制 • 上一篇    下一篇

基于SiamBAN跟踪器改进的目标跟踪算法

郑艳, 赵佳旭, 边杰   

  1. (东北大学 信息科学与工程学院, 辽宁 沈阳110819)
  • 发布日期:2023-09-28
  • 通讯作者: 郑艳
  • 作者简介:郑艳(1963-),女,辽宁沈阳人,东北大学副教授,博士.
  • 基金资助:
    国家自然科学基金资助项目(61773108).

Improved Object Tracking Algorithm Based on SiamBAN Tracker

ZHENG Yan, ZHAO Jia-xu, BIAN Jie   

  1. School of Information Science & Engineering, Northeastern University, Shenyang 110819, China.
  • Published:2023-09-28
  • Contact: ZHAO Jia-xu
  • About author:-
  • Supported by:
    -

摘要: 孪生网络系列的跟踪器基于相似度匹配的方法来实现目标跟踪,当遇到相似干扰物时会发生跟踪漂移现象,从而导致跟踪失败.针对这个问题,以SiamBAN跟踪器为研究基础,提出了一种改进算法.主要改进包括:在训练阶段,加入中心回归分支来降低远离目标中心的边界框分数,同时引入Focal Loss损失函数,在推理阶段设计了全新的筛选策略,来区分要跟踪的目标和相似干扰物.改进后的算法在OTB100测试集的成功率和精度相比于原来分别提高了2.1%和3%,在GOT10k的测试集上成功率比原来提高了2.1%.

关键词: 目标跟踪;SiamBAN;孪生网络;干扰物感知;神经网络

Abstract: The siamese network series tracker utilizes the similarity matching method for object tracking, but tracking drift can occur when similar distractors are encountered, leading to tracking failure. To solve this problem, based on the research of SiamBAN tracker, an improved algorithm is proposed. Major improvements include the addition of a centerness branch during training to reduce bounding box scores far from the object center, the introduction of the Focal Loss function, and a new screening strategy during inference to differentiate the target from similar distractors. Compared with the original, the success plot and precision plot of the improved algorithm are increased by 2.1% and 3% respectively on the OTB100 test set, and the success plot is 2.1% higher than the original on the GOT10k test set.

Key words: object tracking; SiamBAN; siamese network; distractor aware; neural network

中图分类号: