Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (6): 200-207.DOI: 10.3778/j.issn.1002-8331.2010-0149

• Graphics and Image Processing • Previous Articles     Next Articles

Adaptive Decontamination Algorithm Based on PSR Sample Classification

SHI Siqi, MA Yanjun, LI Nanting, ZHENG Liping   

  1. 1.School of Automatic and Information Engineering, Xi’an University of Technology, Xi’an 710048, China
    2.School of Electrical Engineering, Xi’an University of Technology, Xi’an 710048, China
    3.Key Laboratory of Shaanxi Province for Complex System Control and Intelligent Information Processing, Xi’an 710048, China
  • Online:2022-03-15 Published:2022-03-15



  1. 1.西安理工大学 自动化与信息工程学院,西安 710048
    2.西安理工大学 电气工程学院,西安 710048
    3.陕西省复杂系统控制与智能信息处理重点实验室,西安 710048

Abstract: To solve those problems in object tracking such as model drifting and target losing caused by contaminated samples, an adaptive decontamination algorithm based on the peak sidelobe ratio(PSR) sample classification is proposed. This paper analyzes the relationship among the contaminated feature of samples, the response graph of correlation filters and the result of object tracking, and discusses the classification mechanism of samples and the dynamic updating strategy of filter parameters. Different samplesets corresponding to various contamination levels, which are obtained by dividing all samples with the classification thresholds, are respectively used to train those corresponding filters. The weights of some specific sample sets and coefficients of corresponding filters, which are specifically selected with PSR, are dynamically updated. All those correlation filters above are integrated to achieve object tracking. Extensive experiments on OTB-50 and TC-128 verify that the effectiveness in suppressing the affect of severely contaminated samples and improving the accuracy and robustness of object tracking in complex scenes.

Key words: object tracking, correlation filters, adaptive decontamination, sample classification

摘要: 针对相关滤波跟踪算法中训练样本污染容易导致模型漂移和目标丢失等问题,提出基于峰值旁瓣比样本分类的自适应去污算法,建立了样本污染特征、滤波器响应图和跟踪结果之间的内在联系,研究了样本分类机制和参数动态更新策略。通过样本分类阈值将训练样本划分为反映不同污染程度的样本集,并分别训练出对应各类样本集的滤波器;根据峰值旁瓣比动态更新特定样本集的样本权值及其滤波器参数;对各类样本集的相关滤波器进行加权融合实现目标跟踪。在OTB-50和TC-128数据集的测试结果表明:该方法有效抑制了严重污染样本的影响并提高了复杂场景目标跟踪的准确性和鲁棒性。

关键词: 目标跟踪, 相关滤波, 自适应去污, 样本分类