利用自适应频域滤波和凝聚损失的目标跟踪  

Object Tracking Via Adaptive Frequency Domain Filter and Condensation Loss

作　　者：孙培盛 樊佳庆 宋慧慧 SUN Peisheng;FAN Jiaqing;SONG Huihui(Jiangsu Key Laboratory of Big Data Analysis Technology,Collaborative Innovation Center on Atmospheric Environment and Equipment Technology,Nanjing University of Information Science and Technology,Nanjing 210044)

机构地区：[1]南京信息工程大学大气环境与装备技术协同创新中心江苏省大数据分析技术重点实验室,南京210044

出　　处：《计算机与数字工程》2025年第3期725-733,共9页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61532009);江苏省自然科学基金项目(编号:BK20191397)资助。

摘　　要：视觉目标跟踪是计算机视觉领域的一项基础任务,在实际生活中有着广泛的应用,因此视觉目标跟踪技术的研究具有重要意义。之前的跟踪算法存在三个尚未解决的问题,一是随着ResNet网络的提出,空间特征提取得到了大幅加强,但仍有较大的进步空间;二是对于提取出的特征,跟踪器如何突出有效信息、抑制无用信息来达到更好的跟踪效果;三是样本不平衡问题导致模型无法判别未出现过的样本,进而限制了模型的性能。为了解决上述问题,论文设计了双重特征增强模块来优化ResNet骨干网络提取的空间信息。再利用傅里叶变换将空域特征转化为频域特征,接着在频域中利用可学习的滤波器自动过滤出适合当前跟踪器的特征,同时抑制背景中其他杂乱的特征信息。最后,引入凝聚损失函数,扩大困难样本范围并惩罚简单样本,显著缓解了样本不平衡问题。在四个具有挑战性的数据集上的大量实验结果表明,对比同期最先进的算法,论文提出的方法具有更加良好的性能。特别地,在TLP数据集实现了62.0%的成功率得分,在VOT2020LT数据集实现了1.7%的提升。Visual object tracking(VOT)is a fundamental task in computer vision,which has widely applied in realistic scenarios.Therefore,the research on visual object tracking is of great significance.There are three unsolved problems in previous VOT algorithms.Firstly,with the proposal of ResNet,spatial feature extraction is greatly strengthened,but there inherently have several drawbacks.Secondly,for the extracted features,how to leverage useful information and suppress useless information is still challenging for VOT.Thirdly,the sample imbalance leads to the problem that the model lacks the discriminative ability,which limits the performance of the learned model.In order to solve the issues above,this paper uses the dual feature enhancement module to enhance the spatial features extracted by ResNet.The spatial domain features are transformed into frequency domain features by the Fourier transform,and then the suitable features for the tracker are automatically highlighted by employing the learnable filter in the frequency domain,while suppressing other chaotic background information.Finally,the aggregation loss is introduced to expand the range of difficult samples and punish simple samples,which significantly alleviates the problem of sample imbalance.A large number of experimental results on four challenging datasets show that the proposed method has favorable performance against SOTA algorithms.In particular,the proposed approach achieves a AUC score of 62.0%in the TLP and 1.7%improvement in the VOT2020LT.

关 键 词：目标跟踪 双重特征增强模块 自适应频域滤波器 样本不平衡 凝聚损失 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]