基于时空显著性的双核KCF目标跟踪算法

传统的核相关滤波器跟踪算法(KCF)在模板更新上容易出现跟踪误差累计,从而导致目标跟踪过程中出现跟踪漂移问题。针对该问题,提出了一种时空显著性的双核KCF目标跟踪的方法。该算法引入了一种时空显著性方法来搜索目标区域的显著特征和姿态稳定的局部区域。利用该局部区域对跟踪过程中产生的累计误差有较低的敏感度特性,能够减少跟踪过程中的累计误差。然后再结合原目标和显著区域建立一个双核跟踪机制,在跟踪过程中不断对原目标跟踪结果进行微调,降低跟踪累计误差。此外,针对快速运动的目标相邻帧偏移量较大的问题,提出了一种锚点预测机制,使得跟踪锚点与目标位置更接近,能够更准确地跟踪到目标。在大型公共数据上测试的实验结果表明,提出的算法在光照、遮挡、变形、快速运动、旋转以及背景杂波等复杂情况下,均具有较强的适应性。

Dual-core KCF target tracking algorithm based on spatio-temporal saliency

The traditional KCF tracking algorithm is prone to the accumulation of tracking errors in template update,which leads to the tracking drift in the target tracking process.Aiming at this problem,this paper proposed a spatio-temporal significance dual-core KCF target tracking method.This algorithm introduced a spatio-temporal saliency method to search for the significant features of the target region and the local region with stable attitude.The local region was less sensitive to the cumulative errors in the tracking process,which could reduce the cumulative errors in the tracking process.Then,it established dual-core tracking mechanism combining source target and salient region.During the tracking process,the tracking results of the original target are constantly fine-tuned to reduce the cumulative tracking errors.In addition,aiming at the large offset of adjacent frames of fast-moving targets,this paper proposed an anchor point prediction mechanism,which made the tracking anchor closer to the target position and could track the target more accurately.The experimental results on large public data show that the proposed algorithm has strong adaptability in complex situations such as illumination,occlusion,deformation,fast motion,rotation and background clutter.