Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (3): 235-241.DOI: 10.3778/j.issn.1002-8331.2009-0121

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Siamese Network Tracking with Short-Term Memory

WANG Xipeng, LI Yong, LI Zhi, LIANG Qiming   

  1. School of Information Engineering, Engineering University of People’s Armed Police, Xi’an 710086, China
  • Online:2022-02-01 Published:2022-01-28



  1. 武警工程大学 信息工程学院,西安 710086

Abstract: In order to improve the robustness of the target tracking algorithm in complex scenes and adapt the algorithm to long-term tracking scenarios, a tracking method combining multi-layer feature fusion and short-term memory mechanism is proposed to improve the robustness of target tracking. Firstly, the multi-layer features of the convolutional neural network are fused to improve the feature extraction capability of network. Then, in the tracking stage, a short-term memory module is introduced, and the features of the search area are matched with the initial template features and the dynamic features of the short-term memory, and then the response map is fused to improve the robustness of target tracking. The discriminability of the tracking target is enhanced through local information of video. Experiments are carried out on the OTB2015 and GOT-10K target tracking standard data sets. The accuracy and success rate of OTB2015 reach 0.808 and 0.593 respectively. Experimental results show that test effect of the proposed algorithm is significantly improved incomparison with several mainstream tracking algorithms, and the real-time tracking speed reaches 27 frame/s, which proves the effectiveness of the proposed method.

Key words: siamese network, long-term tracking, feature fusion, template update

摘要: 为提升目标跟踪算法在复杂场景下的鲁棒性,使算法适应长时跟踪场景,提出结合多层特征融合和短时记忆机制的跟踪方法,提升目标跟踪的鲁棒性。融合卷积神经网络多层特征,提升网络的特征提取能力。在跟踪阶段,引入了短时记忆模块,搜索区域特征分别与初始的基准模板特征和短时记忆的动态特征进行匹配,对得到的响应图进行融合,提升目标跟踪的鲁棒性,通过视频局部信息增强算法对跟踪目标的判别性。在OTB2015和GOT-10K目标跟踪标准数据集上进行了实验,在OTB2015上的精确度和成功率分别达到了0.808和0.593。实验结果表明,所提算法的测试效果与几种主流跟踪算法相比有了显著的提升,并且达到了27 帧/s的实时跟踪速度,证明了所提方法的有效性。

关键词: 孪生网络, 长时跟踪, 特征融合, 模板更新