引入全局上下文特征模块的DenseNet孪生网络目标跟踪  

DenseNet-siamese Network with Global Context Feature Module for Object Tracking

作　　者：谭建豪[1,2] 殷旺 刘力铭 王耀南 TAN Jianhao;YIN Wang;LIU Liming;WANG Yaonan(College of Electrical and Information Engineering,Hunan University,Changsha 410082,China;National Engineering Laboratory for Robot Visual Perception and Control Technology,Hunan University,Changsha 410082,China)

机构地区：[1]湖南大学电气与信息工程学院,长沙410082 [2]机器人视觉感知与控制技术国家工程实验室,长沙410082

出　　处：《电子与信息学报》2021年第1期179-186,共8页Journal of Electronics & Information Technology

摘　　要：近年来,采用孪生网络提取深度特征的方法由于其较好的跟踪精度和速度,成为目标跟踪领域的研究热点之一,但传统的孪生网络并未提取目标较深层特征来保持泛化性能,并且大多数孪生网络只提取局部领域特征,这使得模型对于外观变化是非鲁棒和局部的。针对此,该文提出一种引入全局上下文特征模块的DenseNet孪生网络目标跟踪算法。该文创新性地将DenseNet网络作为孪生网络骨干,采用一种新的密集型特征重用连接网络设计方案,在构建更深层网络的同时减少了层之间的参数量,提高了算法的性能,此外,为应对目标跟踪过程中的外观变化,该文将全局上下文特征模块(GC-Model)嵌入孪生网络分支,提升算法跟踪精度。在VOT2017和OTB50数据集上的实验结果表明,与当前较为主流的算法相比,该文算法在跟踪精度和鲁棒性上有明显优势,在尺度变化、低分辨率、遮挡等情况下具有良好的跟踪效果,且达到实时跟踪要求。In recent years, the method of extracting depth features from siamese networks has become one of the hotspots in visual tracking because of its balanced in accuracy and speed. However, the traditional siamese network does not extract the deeper features of the target to maintain generalization performance, and most siamese architecture networks usually process one local neighborhood at a time, which makes the appearance model local and non-robust to appearance changes. In view of this problem, a densenet-siamese network with global context feature module for object tracking algorithm is proposed. This paper innovatively takes densenet network as the backbone of siamese network, adopts a new design scheme of dense feature reuse connection network, which reduces the parameters between layers while constructing deeper network, and enhances the generalization performance of the algorithm. In addition, in order to cope with the appearance changes in the process of object tracking, the Global Context feature Module(GC-Model) is embedded in the siamese network branches to improve the tracking accuracy. The experimental results on the VOT2017 and OTB50 datasets show that comparing with the current mainstream tracking algorithms, the Tracker has obvious advantages in tracking accuracy and robustness, and has good tracking effect in scale change, low resolution, occlusion and so on.

关 键 词：目标跟踪 孪生网络 全局上下文特征 DenseNet网络 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]