结合双注意力与特征融合的孪生网络目标跟踪  

Siamese Network Target Trackingthat Combines Dual Attention and Feature Fusion

作　　者：李雪 李晓艳[1] 王鹏[1] 孙梦宇 吕志刚[1] LI Xue;LI Xiaoyan;WANG Peng;SUN Mengyu;LüZhigang(School of Electronic and Information Engineering,Xi'an Technological University,Xi'an 710021,China)

机构地区：[1]西安工业大学电子信息工程学院,西安710021

出　　处：《北京邮电大学学报》2022年第4期116-122,共7页Journal of Beijing University of Posts and Telecommunications

基　　金：国家自然科学基金项目(62171360);陕西省重点研发计划项目(2022GY-110);西安工业大学校长基金面上培育项目(XGPY200217)。

摘　　要：基于孪生网络视觉跟踪的进化和深层网络目标跟踪算法在目标被遮挡和外观形变时的跟踪成功率不高,鲁棒性不强,对此,提出了一种结合双注意力与特征融合的孪生网络目标跟踪算法。首先,采用通道和空间注意力模块增强目标信息,抑制图像中的干扰信息,提高模型的准确度;然后,对注意力层输出的浅层和深层特征信息进行多层特征融合,得到表现力更好的目标特征,提高跟踪成功率;最后,引入在线模板更新机制,减少了跟踪漂移,提高了跟踪鲁棒性。使用OTB100测试集进行实验,实验结果表明,改进后算法的跟踪成功率比改进前算法的跟踪成功率提高了1.3%;在具有遮挡和形变属性的4个测试序列下,改进后算法的平均重叠率提高了3%,中心位置的平均误差降低了0.37个像素点,针对遮挡和外观形变时的鲁棒性更好。In order to solve the problem that the evolution of siamese visual tracking with very deep networks(SiamRPN++) algorithm when the target is occluded and deformed. A siamese network target tracking algorithm combining dual attention and feature fusion is proposed. Firstly, the channel and spatial attention module are used to enhance the target information, suppress the interference information in the image and improve the accuracy of the model;then, multi-layer feature fusion is carried out for the shallow and deep feature information output from the attention layer to obtain better expressive target features and improve the tracking accuracy;finally, the online template update mechanism is introduced to reduce tracking drift and improve the tracking robustness. The OTB100 dataset is used for experimental and the results show that the success rate of the improved algorithm is increased by 1.3% compared with SiamRPN++, indicating that the tracking accuracy of the algorithm is higher;under the four test sequences with occlusion and deformation attributes, the average overlap rate of the improved algorithm is increased by 3%, and the average center position error is reduced by 0.37 pixels, which is better robustness against occlusion and appearance deformation.

关 键 词：目标跟踪 孪生网络 双注意力模块 特征融合 在线模板更新 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]