基于图网络与IoU感知的孪生网络目标跟踪算法
Siamese network object tracking algorithm based on graph network and IoU-aware
陈志旺 1刁华康 2袁宇 3吕昌昊 4彭勇5
作者信息
- 1. 燕山大学智能控制系统与智能装备教育部工程研究中心,河北秦皇岛 066004;燕山大学工业计算机控制工程河北省重点实验室,河北秦皇岛 066004
- 2. 燕山大学工业计算机控制工程河北省重点实验室,河北秦皇岛 066004;航空工业天津航空机电有限公司,天津 300308
- 3. 燕山大学工业计算机控制工程河北省重点实验室,河北秦皇岛 066004
- 4. 燕山大学电力电子节能与传动控制河北省重点实验室,河北秦皇岛 066004
- 5. 燕山大学电气工程学院,河北秦皇岛 066004
- 折叠
摘要
传统孪生网络目标跟踪算法采用互相关或者深度互相关的方式对模板帧与检测帧进行相似性度量,无法有效适应极端的目标形变.以无锚点框的目标跟踪算法为基础,设计一种基于图网络的IoU感知目标跟踪算法.首先,以Resnet50为基础,在每个残差结构后引入通道自适应调整模块NCAM,构造轻量高效且具有通道自适应的特征提取网络.其次,基于图网络设计一种新的模板帧与检测帧的相似性计算方式,将特征图像素点视为图网络的节点,对模板特征与检测特征的图网络节点进行相似性计算以有效应对目标极端形变.最后,在分类部分采用IoU感知的分类损失函数在分类分支与回归分支之间建立联系,改变以往孪生网络目标跟踪算法训练与测试不一致的情况;在回归部分选用CIoU损失计算离线训练阶段的回归损失,得到更加精准的边界框.OTB2015、UAV123、VOT2018与VOT2019数据集上的实验结果验证了所提算法的有效性.
Abstract
Traditional Siamese network target tracking algorithms use cross-correlation or deep cross-correlation to calculate the similarity measurement between template frame and detection frame,which can't effectively adapt to extreme target deformation.In this paper,an IoU-aware target tracking algorithm based on graph network is designed based on the target tracking algorithm without anchor point frame.Firstly,based on Resnet50,the normalization-based channel attention module(NCAM)is introduced after each residual structure to construct a lightweight and efficient feature extraction network with channel adaptive.Then,a new similarity calculation method of template frame and detection frame is designed based on the graph network.The pixel of the feature map is regarded as the node of the graph network,and the similarity calculation is carried out on the node of the graph network with template feature and detection feature to effectively deal with the extreme deformation of the target.Finally,in the classification and regression part,the IoU-perceived classification loss function is used in the classification part to establish a connection between the classification branch and the regression branch,which changes the inconsistency between the training and testing of the Siamese network target tracking algorithm in the past.In the regression part,CIoU loss is used to calculate the regression loss in the off-line training stage,and a more accurate boundary box is obtained.Experimental results on the OTB2015,UAV123,VOT2018 and VOT2019 data sets demonstrate the effectiveness of the proposed algorithm.
关键词
目标跟踪/孪生网络/图网络/IoU感知/通道自适应调整/CIoUKey words
object tracking/Siamese network/graph network/IoU-aware/channel adaptive adjustment/CIoU引用本文复制引用
基金项目
国家自然科学基金项目(61573305)
河北省自然科学基金项目(F2022203038)
河北省自然科学基金项目(F2019203511)
出版年
2024
NETL
NSTL
万方数据