自适应模板更新和目标重定位的相关滤波器跟踪

Correlation filter tracking based on adaptive learning rate and location refiner

针对核相关滤波器在跟踪中因目标快速运动导致的目标易丢失和部分遮挡问题,本文在多特征尺度自适应核相关滤波器(Scale Adaptive with Multiple Features tracker,SAMF)基础上,提出一种融合自适应模板更新和预测目标位置重定位的核相关跟踪算法。采用联合目标移动速度和特征变化的模板更新机制增大对目标快速运动适应性,根据长时滤波器和短时滤波器协作跟踪提出目标位置修正和重定位模型提升跟踪器应对目标部分遮挡的能力。在OTB-2015视频序列集100组序列中与序列集提供的算法进行对比,本算法跟踪精度相比SAMF提升2%。在目标发生快速移动时本文算法具有更好的追踪目标能力,目标重定位也很好地解决了目标部分遮挡问题。

To overcome the problem of loss of target caused by fast motion and the issue of partial occlusion in the tracking of kernel correlation filters,this paper proposed a new kernel correlation tracking algorithm that combines adaptive template updating and the prediction of the relocation of a target,based on the scale adaptive with multiple features tracker(SAMF).A template updating mechanism that combines target velocity and feature changes was proposed to improve the adaptability to fast movement of the target.Based on cooperative tracking of long time and short time filters,a target position correction and relocation model was proposed to improve the ability of the tracker to cope with partial occlusion of the target.In 100 sequences of OTB-2015 video set,the proposed algorithm was compared with the algorithms based on sequence sets and the SAMF algorithm.The tracking accuracy of the proposed algorithm is 2%higher than that of the SAMF algorithm,and the success rate is increased by 1%.The proposed algorithm has better tracking ability for fast moving targets and the target relocation scheme effectively addresses the problem of partial occlusion of the target.