基于互注意力指导的孪生跟踪算法  

Cross-attention Guided Siamese Network Object Tracking Algorithm

作　　者：赵越[1] 余志斌[1] 李永春 ZHAO Yue;YU Zhi-bin;LI Yong-chun(College of Electronic Engineering,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学电气学院,成都611756

出　　处：《计算机科学》2022年第3期163-169,共7页Computer Science

基　　金：装备预研领域基金(61403120304)。

摘　　要：针对传统孪生网络目标跟踪算法在相似物干扰、目标形变、复杂背景等跟踪环境下无法进行鲁棒跟踪的问题,提出了注意力机制指导的孪生网络目标跟踪方法,以弥补传统孪生跟踪方法存在的性能缺陷。首先,利用卷积神经网络ResNet50的不同网络层来提取多分辨率的目标特征,并设计互注意力模块使模板分支与搜索分支之间的信息能够相互流动。然后,在分类与回归网络中,将主干网络提取的每块特征信息权重参数通过神经网络自动学习、更新并加权融合每块特征的分类与回归信息。最后,根据响应图的峰值位置计算目标的预估位置和尺度信息。在UAV123数据集上,所提算法相比主流跟踪算法Siam-BAN,准确率提升了1.7个点,成功率提升了0.7个点;在VOT2018数据集上,相比SiamRPN++算法,所提算法在EAO指标上提高了2.5个点,实时跟踪速度保持在35FPS。Most traditional Siamese trackers cannot perform robust when facing the similar object,deformation,background clutters and other challenges.Accordingly,a cross-attention guided Siamese network(called SiamCAN)is proposed to solve the above problem in this paper.Firstly,different layers of ResNet50are used to get various revolutions of object feature and a crossattention module is designed to bridge the information flow between search branch and template branch.After that,each feature from different layers of backbone is sent to CNNs to update parameters and combined with each other,in classification network and regression network.Finally,the predicted location and target size are calculated according to the max response on response map.Simulation experimental results on the UAV123tracking dataset show that the tracking precision is improved by 1.7%and the tracking accuracy is improved by 0.7%,compared to the mainstream algorithm SiamBAN.Moreover,on the VOT2018benchmark,the EAO of our method outperforms 2.5than the mainstream algorithm SiamRPN++,and the tracking speed of our method maintains 35FPS.

关 键 词：目标跟踪 孪生网络 相似物干扰 互注意力模块 无锚框回归 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]