基于多模型蒸馏的时间正则化相关滤波跟踪算法

目前大多数基于相关滤波的跟踪方法是通过对模型采取简单的线性加权融合或是将历史模型作为时间正则化项来约束模型更新的方式, 增强滤波器对目标的判别能力, 但这种方式对目标时域信息利用有限, 容易造成模型退化漂移。本文提出一种基于多模型蒸馏的时间正则化相关滤波跟踪算法, 该方法通过收集跟踪过程中利用当前样本产生的独立模型, 在建立包含背景信息的局部样本库中来指导滤波器更新, 以此保留目标在时域中的鲁棒特征。同时，根据每一个模型对当前目标的表征能力不同进行可靠性权值更新。最后，利用交替方向乘子(alternating direction multiplier，ADMM)算法进行模型迭代优化。通过在大量的数据库进行实验, 结果表明本文的方法在精确度与成功率上有了大幅提升。

Temporal regularized correlation filter tracking algorithm based on multi-model distillation

At present, most correlation filter based tracking methods adopts simple linear weighted fusion of the model or use the historical model as the temporal regularization term to constrain the model update, which can enhance the ability of the filter to discriminate the target. However, this method cannot make full use of the information of the target, which is easy to cause model degradation and drift. This paper proposes a temporal regularized correlation filter based on multi-model distillation for visual tracking. This method collects the independent model generated by the current sample in the tracking process, which can guide the filter update in the local sample library containing background information. This can retain the robust features of the target in the temporal domain. At the same time, the reliability weight is updated according to the different characterization ability of each model for the current target. Finally, the alternating direction multiplier (ADMM) algorithm is used to iteratively optimize the model. A large number of experimental results in the databases show that the precision and success rate of the method have been greatly improved.