注意力机制的多尺度单目标跟踪算法  

Multi-scale single object tracking based on the attention mechanism

作　　者：宋建锋[1] 苗启广[1] 王崇晓 徐浩 杨瑾 SONG Jianfeng;MIAO Qiguang;WANG Chongxiao;XU Hao;YANG Jin(School of Computer Science and Technology,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学计算机科学与技术学院,陕西西安710071

出　　处：《西安电子科技大学学报》2021年第5期110-116,共7页Journal of Xidian University

基　　金：国家重点研发计划(2019YFC0000238)。

摘　　要：在单目标跟踪过程中,由于存在目标遮挡、目标消失、相似目标干扰等问题,导致算法错误跟踪目标,跟踪精度下降,并且错误的结果将会参与到模型更新中,使得跟踪精度进一步下降。针对这一问题,提出了基于注意力机制的多尺度单目标跟踪算法。该算法使用Inception网络非对称卷积思想,在增加多尺度卷积核的同时减少参数量,非对称卷积可以有效地结合局部特征和全局特征,提高跟踪的鲁棒性。在模型参数更新阶段,采用基于注意力机制的网络在线更新算法,结合每一帧的结果响应图和注意力响应图计算得到该帧的跟踪结果得分,从而在模型更新时剔除不包含目标的视频帧,强化了网络对目标和背景的判别能力,使网络能够快速学习到目标的外观变化,提高算法对目标的跟踪能力。在实验部分,将该算法在OTB-100数据集与其他先进的跟踪算法进行对比,在准确率和成功率方面相较ATOM算法分别有0.9%和0.8%的提升,提升精度的同时也更容易找回丢失的目标。In the process of single object tracking,due to problems of occlusion,disappearance and similar target interference,the tracking accuracy of the algorithm will be reduced.In order to solve these problems,a multi-scale single target tracking algorism based on the attention mechanism is proposed which uses asymmetric convolution to extract the multi-scale feature while reducing the parameters.It combines local features and global features to improve tracking robustness.The online update algorithm based on the attention mechanism is used which combines the response diagram and attention diagram to calculate a score which is used to weed out frames without targets.The attention mechanism strengthens the ability to distinguish the target and background,makes the network quickly adapt to the changes of the appearance of the targets,and improves the tracking performance of the algorithm.The algorithm is tested on OTB-100 datasets with other advanced tracking methods.Compared to the ATOM,the accuracy and success rate of our method are improved by 0.9%and 0.8%respectively and it is easier to retrieve the target after it is lost.

关 键 词：深度学习 单目标跟踪 注意力机制 卷积神经网络 

分 类 号：TP37[自动化与计算机技术—计算机系统结构]