具有改进时空正则和异常抑制的相关滤波跟踪

Correlation filter tracking with improved spatio-temporal regularization and abnormal suppression

为提高正则化相关滤波跟踪算法在复杂场景下的跟踪性能,提出一种具有改进的时空正则化和异常抑制的相关滤波目标跟踪算法。该算法在现有时空正则项的基础上引入稀疏空间变化向量来学习空间权重和参考权重之间的变化,引入稀疏时间变化向量来学习相邻两帧之间滤波器的变化,同时将响应图变化引入目标函数来抑制训练过程中的响应突变,最后通过交替方向乘子法迭代优化求解。在OTB-2015数据集上的对比实验结果表明,提出的算法具有更好的精度和成功率,且在光照变化、遮挡、快速运动等复杂跟踪场景下具有更好的鲁棒性。

To improve the tracking performance of regularized correlation filter tracking algorithms in complex tracking scenarios, a correlation filter tracking algorithm with improved spatio-temporal regularization and abnormal suppression was proposed. Based on the existing spatio-temporal regularization term, a sparse spatial variation vector was introduced to learn the changes between spatial weight and reference weight, and a sparse temporal variation vector was introduced to learn the changes of filters between two adjacent frames, meanwhile, the change of response map was added into the objective function to suppress the sudden change of response during the training process. Finally, the iterative optimization is solved by alternating direction multiplier method. The comparative experimental results on OTB-2015 dataset show that the proposed algorithm has better precision and success rate, as well as better robustness in various complex tracking scenarios such as illumination variation, occlusion and fast motion.