摘要
基于相关滤波理论的判别式跟踪方法由于其高效性和鲁棒性,已经取得了一系列的进展,成为了目标跟踪领域的研究热门.为了使更多国内外学者对相关滤波目标跟踪理论及其发展进行进一步研究与探索,对该领域研究现状进行综述.首先,介绍了相关滤波理论及其用于实现目标跟踪任务时的一般框架,并重点描述了典型的核相关滤波跟踪方法.其次,讨论了目标跟踪技术应用于实际场景时面临的诸多难题,详细分析了特征表示和自适应尺度更新这2个主要难点.然后,从基本类相关滤波、部件类相关滤波、正则化类相关滤波和Siamese网络类相关滤波这4个类别对具有代表性的算法进行分析与讨论,并指出了未来可能的发展趋势.最后,在OTB2013和OTB100基准数据集上对32种相关滤波类跟踪算法就精确度、成功率和帧率进行了对比,在VOT2017数据集上对10种相关滤波类跟踪算法就平均重叠期望(expected average overlap,EAO)、Accuracy和Robustness三个性能指标进行了对比,体现了相关滤波跟踪器(correlation filter trackers,CFTs)的优越性.尽管相关滤波理论在目标跟踪领域具有广阔的应用前景,但是复杂场景和自身因素的影响导致其仍然是一个极具挑战性的研究方向,研究兼备准确性与鲁棒性的CFTs对于目标跟踪领域的发展具有重要意义.
In the field of object tracking,the discriminant method based on correlation filter theory has made a series of advances and becomes a hot research topic because of its efficiency and robustness.The current research statuses of the tracking field were reviewed to allow more scholars to explore the theory and development of correlation filter-based trackers.First,the correlation filter theory and the general framework for object tracking were introduced,and the classical kernelized correlation filter was described in detail.Second,the difficulties of object tracking technology when applied in the real application were discussed,and the main difficulties of feature representation and adaptive scale updating were analyzed in detail.Then,the representative algorithms were analyzed and discussed from the four categories of basic correlation filter,part correlation filter,regularized correlation filter,and Siamese network correlation filter,and the possible future development trend was pointed out.Finally,32 types of correlation filter-based trackers were compared in terms of accuracy,success rate and frame rate on the OTB2013 and OTB100 standard data sets,and 10 types of correlation filter-based trackers were compared in terms of EAO,and the accuracy and robustness on the VOT2017 data set,further indicated the advantages of correlation filter tracking algorithms.The research on correlation filter theory has extensive applications in the object tracking field.However,it is still a challenging research direction due to the influence of complex scenes and their own factors.Developing a highly efficient and robust correlation filter tracker is considered significant.
作者
孟晓燕
段建民
MENG Xiaoyan;DUAN Jianmin(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)
出处
《北京工业大学学报》
CAS
CSCD
北大核心
2020年第12期1393-1416,共24页
Journal of Beijing University of Technology
基金
北京市自然科学基金资助项目(JJ002790200802)
北京市属高等学校人才强教计划资助项目(038000543117004)。
关键词
机器视觉
目标跟踪
相关滤波
特征表示
尺度更新
孪生网络
性能指标
computer vision
object tracking
correlation filter
feature representation
scale updating
siamese network
performance indicator