SiamCMN: Jointing Cycle Memory and Correlation Network for Siamese Object Tracking
SiamCMN: Jointing Cycle Memory and Correlation Network for Siamese Object Tracking
复制标题
DOI：
10.1016/j.ijleo.2023.170981
复制
发表时间：
2023-05
期刊：
影响因子：
3.1
通讯作者：
Baohua Zhang;Nianchao Zhang;Yongxiang Li;Xiaoqi Lu;Yu Gu;Jian Li
中科院分区：
文献类型：
--
作者：
Baohua Zhang;Nianchao Zhang;Yongxiang Li;Xiaoqi Lu;Yu Gu;Jian Li
在具有潜在非确定性状态的异构场景中，模板特征决定了基于连体网络的跟踪器的性能，然而，在搜索区域匹配过程中很容易引入背景噪声，这导致跟踪器缺乏足够的辨别力。针对这一问题，本文提出一种基于循环存储模块的目标跟踪算法。首先，将提取的主要特征映射到循环空间存储网络，空间存储单元用于存储干扰分量。在此基础上，利用循环结构滤除干扰成分，细化模板特征。然后，该方法引入了通道相关网络，该网络可以通过对通道之间的正相关和负相关进行建模来挖掘语义互补信息并捕获全局上下文相关性。在 OTB50、OTB100、UAV123、VOT2018、GOT10k 和 LaSOT 等跟踪基准上的实验结果表明，所提出的方法在挑战性场景中的跟踪精度优于最新的代表性算法。
In heterogeneous scenes with latent non-deterministic states, template features determine the performance of the Siamese network-based trackers, however, background noise is easily introduced in the search area matching process, which causes the trackers lack sufficient discrimination. To solve this problem, this paper proposes an object tracking algorithm based on the cyclic memory module. Firstly, the extracted primary features are mapped to the cyclic spatial memory network, the spatial memory unit is used to store the interference components. On this basis, the cyclic structure is used to filter out the interference components and refine the template features. Then a channel correlation network is introduced into the proposed method, which can mine semantic complementary information and capture global contextual relevance by modeling the positive and the negative correlation between the channels. The experimental results on tracking benchmarks such as OTB50, OTB100, UAV123, VOT2018, GOT10k, and LaSOT have shown that the proposed method outperforms recent representative algorithms in tracking accuracy in challenging scenarios.