用于目标跟踪的特征融合孪生网络算法研究

Study of the feature fusion siamese network algorithm for target tracking

针对目标跟踪过程存在的动态不确定性的问题,传统跟踪方法容易产生目标漂移甚至跟踪失败,而基于深度学习的跟踪算法随着网络结构的加深容易导致深层特征过于稀疏抽象,不利于克服上述问题.为此,本文提出SiamMask三分支网络融合注意力机制的孪生网络目标跟踪新方法,旨在加强网络对特征选取的学习能力,加强目标有效特征的抽取,并减少冗余信息对网络负担的影响.特征提取主干网络选用改进的Resnet-50,通过融合深层和浅层特征,实现跟踪目标特征的有效表达.利用4个数据集(COCO、ImageNet-DET 2015、ImageNet-VID 2015、YouTube-VOS)对提出的特征融合孪生网络框架进行训练,并使用VOT数据集进行在线测试.实验表明:与文中其他跟踪方法相比,该算法在面对动态目标尺度变化、环境光照、运动模糊等场景表现更优异.

During processes of target tracking,traditional tracking methods tend to face challenges of dynamic uncertainty by inducing target drifts or even complete tracking failures.At the same time,due to the deepening of network structures,features of tracking algorithm based on deep learning become too sparse and abstract,thus leading to difficulties of tackling these problems.Consequently in this study,we propose a new siamese network structure with SiamMask three-branch and fuse attentional mechanism.We aim to strengthen the network learning ability for feature selection,highlight the extraction of target’s effective features,and reduce the impact of redundancy information on the network burden.In this method,we use improved Resnet-50 as a feature extraction backbone network so that the effective expression of tracking target by integrating deep and shallow features can be achieved.The proposed feature fusion siamese network framework is developed by learning four datasets(COCO,ImageNet-DET 2015,ImageNet-VID 2015 and YouTube-VOS)and is tested using the VOT dataset.Experiments show that,compared with other tracking methods in this paper,the present algorithm performs more satisfactorily when facing dynamic targets with scale changes,environment lighting,also motion-ambiguity scenarios.