基于时间正则化及背景感知的滤波器跟踪

Filter Tracking Based on Time Regularization and Background-Aware

针对相关滤波器(CF)的目标背景因没有根据时间建模而导致的性能不佳的问题,在方向梯度直方图(HOG)的基础上,提出一种基于时间正则化及背景感知的滤波器跟踪算法。从真实的背景中提取训练样本,通过增加训练样本来增强滤波器的分类能力;引入时间正则化,构建遮挡情况下目标重定位模块;采用交替方向乘子法(ADMM)优化求解目标,降低计算复杂度;采用线性插值策略来更新目标的位置和尺度。采用目标跟踪基准(OTB-2015)数据集中的100个视频序列与评价标准对本文所提出的算法进行性能测试。实验结果表明,基于时间正则化及背景感知的滤波器跟踪算法的精确度得分达到0.801,成功率得分为0.762,相比核相关滤波器(KCF)算法分别提高了20%和46.8%。本文算法能很好解决目标发生平面外旋转、目标被遮挡、背景嘈杂等情况下的视觉跟踪问题,具有良好的应用前景和较大的使用价值。

This study proposes a filter tracking algorithm based on the direction gradient histogram using time regularization and background-aware to overcome the problem of target background of the correlation filter(CF)having no optimal performance without time modeling.The training samples are firstly extracted from the real background,and classification ability of the filter is enhanced by adding the training samples.Subsequently,time regularization is introduced to construct the target relocation module under occlusion.In addition,the alternating direction multiplier method is used to optimize the solution target and reduce the computational complexity.Finally,a linear interpolation strategy is used to update the target location and scale.The proposed algorithm uses 100 video sequences and evaluation criteria in object tracking benchmark(OTB-2015)dataset for performance testing.Experimental results show that the accuracy score of filter tracking algorithm using time regularization and background-aware reaches 0.801 and success rate score is 0.762,which are 20% and 46.8% higher,respectively,compared to those of the kernelized correlation filter(KCF)algorithm.The proposed algorithm can solve the visualtracking problem of off-plane rotation,occlusion,and background ambiguity,which has wide application prospects and use value.