动态权重层次卷积特征自适应目标追踪算法  

Adaptive Target Tracking Algorithm with Dynamic Weighted Hierarchical Convolutional Neural Network Feature

作　　者：杜丰 王万良[1] 李思远 张智 刘子瑜 DU Feng;WANG Wan-liang;LI Si-yuan;ZHANG Zhi;LIU Zi-yu(School of Computer Science,Zhejiang University of Technology,Hangzhou 310023,China;Department of Computer Science and Technology,Nanjing University,Nanjing 210023,China;School of Information Engineering,Zhijiang College of Zhejiang University of Technology,Shaoxing 312030,China)

机构地区：[1]浙江工业大学计算机学院,杭州310023 [2]南京大学计算机科学与技术系,南京210023 [3]浙江工业大学之江学院信息工程学院,浙江绍兴312030

出　　处：《小型微型计算机系统》2021年第4期835-841,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61873240)资助。

摘　　要：为改进追踪算法在目标快速运动或被遮挡等情况下的性能,对目标追踪中普遍采用的相关滤波算法框架进行了研究,基于核相关滤波器(KCF)提出一种层次化提取卷积神经网络特征并自适应赋予动态权重的目标追踪算法.通过提取不同层次卷积神经网络特征,分别经过相关滤波器学习得到不同的KCF模板,结合特征层次和各滤波器稳定度、准确度赋予动态权重,以融合3个模板确定最终目标位置.实验采用OTB标准数据库,测试了新算法在遮挡、运动模糊、快速运动等干扰项下的整体性能,结果表明所提算法在整体上提高了追踪的性能及精度,可以灵活适应不同特征的场景,并且相较于经典KCF平均精确度提高了35.4%,平均成功率提高了33.6%.In order to improve the performance of the tracking algorithm in the case of fast motion or occlusion,a dynamic weighted hierarchical convolution feature adaptive target tracking algorithm is proposed on the basis of the kernel correlation filter( KCF)framework commonly used in target tracking. After extracting features of different layers of different convolution neural network,different KCF templates were obtained through correlation filter learning. By combining stability and accuracy of each filter and giving dynamic weights to different features,stability and accuracy of each filter,and fusing three templates,the final target position could be determined. The OTB standard database was used to test the performance of the new algorithm under the disturbances such as occlusion,motion blur,fast motion,etc.. The results show that the proposed algorithm improves the tracking performance and accuracy,and can flexibly adapt to scenarios with different features. Compared with the classic KCF,the average accuracy is increased by 35.4%,and the average success rate is increased by 33.6%.

关 键 词：卷积神经网络 核相关滤波 动态权重 层次卷积特征 目标追踪 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]