融合检测机制的鲁棒相关滤波视觉跟踪算法

Fusion detection mechanism of robust correlation filtering visual tracking algorithm

为了解决相关滤波视觉跟踪算法在复杂场景中产生的跟踪漂移问题,提出一种融合检测机制的相关滤波跟踪框架。利用时空正则化滤波器作为跟踪器,同时使用线性核相关滤波器作为检测器。当跟踪器与目标进行相关计算得到的响应图为多个峰值时,激活检测器,对多个峰值进行相关匹配,获得重检测结果;同时,使用平均峰值相关能量的滤波器模型更新策略得到更加可靠的检测器,以达到提高跟踪精度和算法鲁棒性的目的。在OTB2015、Templecolor128和VOT2016数据平台上的实验结果表明,与近年提出的性能较出色的跟踪算法相比,本文算法在目标运动模糊、相似背景干扰和光照变化等复杂场景中具有更好的鲁棒性和准确性,且跟踪精度和成功率上均有提高。

In order to solve the tracking drift problem caused by the correlation filtering visual tracking algorithm in complex scenes,a correlation filtering tracking framework fused with detection mechanism was proposed.A space-time regularization filter was used as a tracker,and a linear kernel correlation filter was used as a detector.When the response diagram obtained by correlating the tracker with the target was a plurality of peaks,the detector was activated to perform correlation matching on multiple peaks to obtain a retest result;meanwhile,a filter model update strategy using average peak correlation energy was used to obtain a more reliable detector,so as to improve the tracking accuracy and algorithm robustness.The experimental results on the OTB2015,Temple color 128 and VOT2016 data platforms show that compared with the tracking algorithms of better performance proposed in recent years,the proposed algorithm has better robustness and accuracy in complex scenes such as target motion blur,similar background interference and illumination changing,and both of the tracking accuracy and the success rate are improved.