高效的多特征自适应相关滤波跟踪器

An Efficient Tracker via Multi-feature Adaptive Correlation Filter

针对基于手工特征的相关滤波跟踪器在快速变形、背景杂乱等挑战性跟踪场景效果不佳的问题,在Staple跟踪器的基础上提出了一种新型的相关滤波跟踪器。结合方向梯度直方图特征与颜色命名特征构建了目标外观模型,增强其对快速变形以及背景杂乱等场景的鲁棒性;设计了自适应评分函数对2种特征进行融合,得到更具有鉴别性的特征;针对不同的特征分别提出了在线更新策略以减小训练过拟合与模型漂移。实验结果表明:该跟踪器在跟踪的准确性与实时性上均有着优良表现。

Aiming at the low tracking effect of the correlation filters tracker based on manual features in challenging scenes of rapid deformation and background clutter,a new correlation filter tracker based on Staple tracker is proposed.An appearance model based on HOG features and color-naming features is built to enhance the robustness to the challenging scenes of rapid deformation and background clutter.A self-adjust evaluation function is designed to merge the two kinds of feature information and a more discriminative feature is obtained.The novel online update strategies to reduce the training over-fitting and model drift for different features are proposed.The tracker shows excellent performance in accuracy and real-time capability on OTB2015 benchmark.