基于深浅层特征融合的孪生网络目标跟踪算法  

Siamese network object tracking algorithm based on deep and shallow feature fusion

作　　者：王子康 姚文进[1] 薛尚捷 司婷波 WANG Zikang;YAO Wenjin;XUE Shangjie;SI Tingbo(School of Mechanical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学机械工程学院,南京210094

出　　处：《北京航空航天大学学报》2025年第3期973-984,共12页Journal of Beijing University of Aeronautics and Astronautics

摘　　要：深度学习相关的目标跟踪算法在利用深浅层特征融合时,未考虑分类分支与回归分支的差异性,两分支均使用同一融合特征,不能同时满足各自分支的不同任务要求。依据分类分支与回归分支的不同任务要求与深浅层特征之间的关系,提出了一种新的特征融合方式用于视觉目标跟踪。将骨干网络中不同特征层的通道数按比例进行微调,分别形成适合分类分支与回归分支的融合特征。为验证所提特征融合方式的有效性,在基于SiamCAR算法的基础上进行优化,改变特征提取与融合方式,在UAV123、GOT-10K、LaSOT数据集上提高了2%~3%的精度。实验结果证明:所提特征融合方式是有效的,同时框架整体以75帧/s的实时运行速率实现了良好的跟踪性能。Until recently,object tracking algorithms connected to deep learning have not considered the distinctions between regression and classification branches while employing shallow and deep feature fusion.Both branches used the same fusion feature,which could not meet the different task requirements of each branch well at the same time.According to the relationship between different task requirements of branches and features,a new feature fusion method was proposed for the object tracking algorithm.The channels of different feature layers in the backbone network were adjusted proportionally to form the fusion features suitable for the classification branch and regression branch.To compare and prove the effectiveness of the new feature fusion method,optimization was carried out on the basis of the SiamCAR algorithm.By changing the method of feature extraction and fusion,the accuracy of 2%−3%was improved on the three datasets of UAV123,GOT-10k and LaSOT.The experimental findings demonstrate the efficacy of the novel feature fusion technique,and the framework as a whole achieves good tracking performance at a real-time running speed of 75 frames per second.

关 键 词：目标跟踪 孪生网络 分类 回归 深浅层特征融合 分类与IoU联合训练 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]