针对相关滤波目标跟踪算法空间正则权重没有与目标建立联系和时间正则项不能自适应更新的问题,提出自适应时空正则化的相关滤波目标跟踪算法。首先,利用初始帧的显著感知参考权重,使自适应空间正则项能够在后续跟踪过程中获取与目标存在联系的空间正则权重。然后,利用相邻两帧响应得分的变化情况计算时间正则化参数的参考值,使自适应时间正则项可以通过变化的正则化参数不断更新。最后,采用交替方向乘子法(ADMM)优化算法,以较少的迭代次数分别求解出滤波器函数、空间正则权重和时间正则化参数。在OTB-2015数据集上进行实验,结果表明本文算法的跟踪性能优于其他对比算法,其中距离精度和成功率分别达到86.4%和65.6%,且本文算法在具有形变、旋转、遮挡和出视野等属性的复杂跟踪场景下更具鲁棒性。

For current correlation filter target tracking algorithm, the spatial regularization weight is not connected with an object, and the temporal regularization term fails to update adaptively. To resolve this problem, a correlation filter based on adaptive spatiotemporal regularization was proposed. The adaptive spatial regularization term first obtains the spatial regularization weight connected with the object by initial-frame saliency aware reference weight. Second, the reference value of the temporal regularization parameter is calculated using the altered response score between two adjacent frames. Thus, the adaptive temporal regularization term can be continuously updated by the changing regularization parameter. Finally, the algorithm is optimized by the alternating direction method of multipliers, which reduces the number of iterations and solves the related parameters (filtering function, spatial regularization weight, and temporal regularization parameter). In an experimental evaluation on OTB-2015 dataset, our algorithm outperformed comparable algorithms, achieving a distance precision of 86.4% and a success rate of 65.6%. The proposed algorithm also showed higher robustness in complex scenes with deformation, rotation, occlusions, and out of view than the competing algorithms.