空间正则化支持相关滤波算法

Spatially regularized support correlation filter algorithm

支持相关滤波跟踪方法利用循环采样将计算转换到频域进行,解决了支持向量机采样少与计算量大的问题。但是当前方法在跟踪过程中对目标周围进行循环采样,产生大量移位样本,不能很好地利用背景信息。针对该问题,提出空间正则化支持相关滤波算法。该算法在训练过程中引入空间正则化分量,根据滤波器的空间位置对其进行惩罚。该算法能够在更大的区域进行采样,使用更多的真实样本进行训练,同时,进一步使用ADMM优化策略提高了跟踪速度。在OTB100数据库上的实验表明,相对于尺度核化支持相关滤波算法(SKSCF),所提算法在精确度和成功率上分别提高了4.3%和11.2%。

In the support correlation filter tracking method,calculations are converted into frequency domain by cyclic sampling,which eliminates less sampling and high computational complexity of support vector machine(SVM).However,circular sampling around the target will generate a large number of shifted samples and the background information cannot be utilized effectively during the tracking process.Therefore,a spatially regularized support correlation filter algorithm is proposed.In this method,the spatially regularized component is introduced in the training process,and the filter is penalized according to its spatial location.The algorithm can sample in a larger area and use more real samples for training.Furthermore,ADMM(alternating direction method of multipliers)optimization strategy is used to improve the tracking speed.Experiments on the database OTB100 show that the proposed algorithm can improve accuracy and success rate respectively by 4.3%and 11.2%in comparison with the scale kernel support correlation filter(SKSCF)algorithm.