摘要
基于孪生网络的深度学习算法是实现目标跟踪的重要手段,但常见的基于孪生网络的跟踪算法存在无法有效利用多尺度信息和历史信息的问题。针对这些问题,本文在SwinTrack网络的基础上进行改进,首先,使用可变形的自注意力模块构建主干网络,以增强特征提取能力;然后,设计一种多尺度特征融合结构,以增加模型对浅层特征和多尺度特征的利用;最后,提出一种模板融合更新结构,构建当前信息与历史信息相融合的单目标跟踪模型,进一步提升跟踪的精度。实验结果表明,本文所提出的目标跟踪网络在GOT-10k数据集上达到了73.2%的准确率,较基准网络在AO指标上提升了4.3百分点,在LaSOT、TrackingNet、LaSO_(Text)数据集上的准确率分别为70.8%、82.9%和48.3%,较基准网络分别提升了2.4百分点、1.9百分点和2.4百分点,且相比于同类型网络有更高的准确率,验证了本文所设计网络的合理性。
Deep learning algorithm based on siamese network is an important means to achieve object tracking,but common tracking algorithms based on siamese network can not effectively use multi-scale information and historical information.To address these problems,this paper makes improvements on the basis of SwinTrack network.Firstly,a deformable self-attention module is used to build a backbone network to enhance feature extraction capability.Then,a multi-scale feature fusion structure is designed to increase the use of shallow features and multi-scale features.Finally,a template fusion update structure is proposed to build a single object tracking model with the fusion of current and historical information,which further improves the tracking accuracy.The experimental results show that the object tracking network proposed in this paper has achieved an accuracy rate of 73.2%on the GOT-10k dataset,which is 4.3 percentage points higher than that of the benchmark network in terms of the AO index.The accuracy rates on the LaSOT,TrackingNet,and LaSO_(Text) datasets are 70.8%,82.9%,and 48.3%respectively,which are 2.4 percentage points,1.9 percentage points and 2.4 percentage points higher than those of the benchmark network.Moreover,compared with similar networks,this network has higher accuracy rates,which verifies the rationality of the network designed in this paper.
作者
吕熊
何月顺
何璘琳
汪显顺
张维
LYU Xiong;HE Yueshun;HE Linlin;WANG Xianshun;ZHANG Wei(East China University of Technology,Nanchang 330013,China)
出处
《计算机与现代化》
2026年第2期120-126,共7页
Computer and Modernization
基金
江西省03专项及5G项目(20232ABC03A09)。
关键词
目标跟踪
孪生网络
多尺度特征
模板更新
注意力机制
object tracking
siamese network
multi-scale feature
template update
attention mechanism
