Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (8): 213-224.DOI: 10.3778/j.issn.1002-8331.2211-0435

• Graphics and Image Processing • Previous Articles     Next Articles

Siamese Networks for Object Tracking on Statistical Characteristics of Distributions

LI Jun, CAO Lin, ZHANG Fan, DU Kangning, GUO Yanan   

  1. 1.School of Instrument Science and Opto-Electronics Engineering, Beijing Information Science and Technology University, Beijing 100101, China
    2.Key Laboratory of the Ministry of Education for Optoelectronic Measurement Technology and Instrument, Beijing Information Science and Technology University, Beijing 100101, China
    3.Key Laboratory of Information and Communication Systems, Ministry of Information Industry, Beijing Information Science and Technology University, Beijing 100101, China
  • Online:2024-04-15 Published:2024-04-15

分布统计特征的孪生网络目标跟踪方法

李俊,曹林,张帆,杜康宁,郭亚男   

  1. 1.北京信息科技大学 仪器科学与光电工程学院,北京 100101
    2.北京信息科技大学 光电测试技术及仪器教育部重点实验室,北京 100101
    3.北京信息科技大学 信息与通信系统信息产业部重点实验室,北京 100101

Abstract: Although siamese trackers have achieved great success, the tracking performance is inferior in complex scenes such as ambiguous boundaries. Most of the existing methods use the inflexible Dirac distribution for target localization. Due to the lack of uncertainty estimation of the bounding box, the target cannot be accurately located under the ambiguous boundaries. For this purpose, this paper improves SiamBAN. Firstly, the representation of the bounding box is changed from Dirac distribution to the general distribution within a certain range with the help of the characteristics that distribution statistics of the bounding box are highly correlated with the actual localization quality. Secondly, a higher localization quality estimation score is generated by putting the distribution statistics into distribution guided quality predictor. Finally, classification and localization quality estimation are represented jointly which can overcome the problem of inconsistency between classification and localization in training and testing stages. Extensive experiments on visual tracking datasets including VOT2018, VOT2019, OTB100, UAV123, LaSOT, TrackingNet, and GOT-10k demonstrate that the performance of proposed method surpass SiamBAN by 3.3%~10% in terms of accuracy and EAO.

Key words: siamese network, localization quality, uncertainty estimation, distribution statistics, distribution guided quality predictor

摘要: 尽管基于孪生网络的跟踪器取得了巨大成功,但在边界模糊这类复杂场景下的跟踪性能仍然较差。大多数现有方法对于目标的定位均采用不灵活的狄克拉分布,由于缺少对边界框的不确定性估计,使其在边界模糊下无法准确定位。为了解决上述问题,基于SiamBAN模型进行改进,利用目标边界框的分布统计特征与其实际的定位质量高度相关这一特性,将边界框的回归值由狄克拉分布转为一定范围内的任意概率分布,将分布统计特征经过分布引导质量预测器生成较高的定位质量估计得分,将分类与定位质量估计联合表示,克服了训练和测试阶段分类与回归不一致问题。在VOT2018、VOT2019、OTB100、UAV123、LaSOT、TrackingNet和GOT-10k数据集上的实验结果表明,对比SiamBAN在准确度和EAO指标上提升了3.3%~10%。

关键词: 孪生网络, 定位质量, 不确定性估计, 分布统计特性, 分布引导质量预测器