计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (20): 206-211.DOI: 10.3778/j.issn.1002-8331.2103-0098

• 图形图像处理 • 上一篇    下一篇

基于原型注意力的多域网络目标跟踪方法

刘满,胡磊,宁纪锋,刘扬   

  1. 西北农林科技大学 信息工程学院,陕西 杨凌 712100
  • 出版日期:2022-10-15 发布日期:2022-10-15

Multi-Domain Network Target Tracking Method Based on Prototype Attention

LIU Man, HU Lei, NING Jifeng, LIU Yang   

  1. College of Information Engineering, Northwest A&F University, Yangling, Shaanxi 712100, China
  • Online:2022-10-15 Published:2022-10-15

摘要: 为了解决预训练集和跟踪视频的域不一致性导致跟踪模型判别能力不足的问题,提出了一种基于原型注意力的多域网络目标跟踪方法。以实时多域网络目标跟踪方法为研究对象,在训练过程中引入原型网络提取注意力特征。基于支撑集正负样本得到目标与背景的域特定原型注意力,将其与待跟踪视频的特征图进行逐通道自适应融合,使得模型在大型数据集上得到判别力更强的目标表示,从而增强跟踪算法的性能。在OTB100和TrackingNet两个基准数据集上的实验结果表明,提出方法的精度和成功率优于现有的代表性跟踪方法。

关键词: 原型网络, 注意力机制, 目标跟踪, 元学习

Abstract: A multi-domain network object tracking method, which based on prototype attention is proposed to solve the problem that the domain inconsistency between the pre-training datasets and the tracking video results in insufficient discriminant ability of the tracking model. It takes the real-time multi-domain network target tracking method as the research object, and introduces the prototype network to extract attention features during training. Based on the positive and negative samples generated by support set, the domain-specific prototype attention of the target and background is obtained, and it performs channel-wise adaptive fusion with the feature map of the video sequence to be tracked, which trains the model to get a more discriminant target representation on a large dataset, which enhances the performance of the tracking algorithm. On the two benchmarks of OTB100 and TrackingNet, the proposed Proto-MDNet tracking method performs better than the existing representative tracking methods in both success rate and precision rate.

Key words: prototype network, attention mechanism, object tracking, meta learning