基于空间正则化约束的支持向量相关滤波器目标跟踪方法
2021-07-11李峰张宏志左旺孟
李峰 张宏志 左旺孟



摘 要:基于支持向量相关滤波器(Support Correlation Filters,SCF)的目标跟踪方法存在严重的样本边界不连续问题,因此模型判别能力受到严重限制。本文将空间正则化项引入到SCF中,提出了基于空间正则化约束的支持向量相关滤波器(Spatially Regularized SCF,SRSCF)模型。相比于SCF,SRSCF不仅可以借助更大的图像区域进行模型学习,同时也能缓解样本的边界不连续问题对模型学习的负面影响,由此得到判别能力更强的模型。此外,本文提出了一种ADMM(Alternating Direction Method of Multiplier)算法求解SRSCF模型,其中每个子问题具有解析解。实验结果表明,相较于SCF,SRSCF能够有效地提升跟踪精度,同时仅增加较少的计算开销。
关键词: 目标跟踪;支持向量相关滤波器;空间正则化
文章编号: 2095-2163(2021)01-0147-05 中图分类号:TP391.41 文献标志码:A
【Abstract】The existing support correlation Filters (SCF) methods suffer from unwanted boundary discontinuity problem of samples, resulting in the degraded CF models. To address this, this paper incorporates the spatial regularization term into the SCF method, and proposes the spatially regularized SCF (dubbed SRSCF) model. In comparison to SCF, SRSCF can leverage larger image regions during model learning, and also alleviate the negative impacts of boundary discontinuous samples on model training, thereby leading to more discriminative CF models. In addition, an ADMM algorithm is proposed to solve the proposed SRSCF model, in which each sub-problem has closed-form solution. Experimental results show that SRSCF can achieve better performance than the SCF models, and only need less additional computational overhead.
【Key words】object tracking; Support Correlation Filter; spatial regularization
0 引 言
近年来,相关滤波器(Correlation Filter,CF)方法凭借其良好的计算效率和稳定的性能在目标跟踪领域取得了巨大的成功。基于CF的跟踪方法旨在学习具有判别能力的二类分类器,通过与样本进行循环卷积运算,以生成预先定义好的标签响应图。早期的CF方法,如MOSSE[1] , KCF[2] 等,将分类器学习表示为l2范数约束下的岭回归模型。由于SVM分类器在小样本学习中拥有良好的性能,Rodriguez等人[3] 将SVM和CF模型相结合,提出最大边界CF分类器。但由于该方法需要提前对样本进行线下训练,因此无法应用于在线的目标跟踪任务。为了解决该问题,Zuo等人[4] 将SVM模型嵌入到CF跟踪框架下,提出了支持向量CF模型—SCF。SCF不仅可以借助快速傅里叶变换加速模型学习,同时也能够灵活地扩展到多通道、核空间中以提升模型的判别能力。……
