基于前景感知的时空相关滤波跟踪算法

Foreground-Aware Based Spatiotemporal Correlation Filter Tracking Algorithm

针对长时目标跟踪中目标背景混杂、遮挡、目标移出视野导致的跟踪失败问题,基于空间正则化相关滤波(SRDCF),提出一个基于前景感知的时空相关滤波算法。首先,提出前景感知相关滤波方法,使得滤波器能够准确地把目标前景区域和背景区域进行区分;然后,把前景感知滤波器加入时间正则项中,使具有时空正则化功能的滤波器始终保持在一个低维的判别流形上;同时,采用交替方向乘子法(ADMM)求解,使得跟踪方法在传统特征的表达上能实现实时性;最后,确定目标重检测器的激活阈值,利用候选区域方法结合相关滤波方法实现重检测,达到长时跟踪的目的。在标准数据集OTB-2013上分别利用传统特征和卷积特征进行实验,并与SRDCF相比,跟踪平均成功率分别提高了5.6%和7%。本文算法针对目标背景模糊、旋转、遮挡和移出视野等情况,具有较强的稳健性。

In this study,we propose a foreground-aware based spatiotemporal correlation filter algorithm based on the spatially regularized discriminative correlation filter(SRDCF)to deal with long-term object tracking failures caused by background clutter,occlusions,and out-of-view objects.Initially,a foreground-aware correlation filtering algorithm is proposed to distinguish the foreground and background of the object accurately.Subsequently,the foreground-aware filter is added to the time regularization term to keep the filter with spatiotemporal regularization function in a low-dimensional discriminative manifold.Simultaneously,the solution based on the alternating direction method of multipliers(ADMM)is conducted to achieve real-time operation of the tracking method in the traditional feature expression.Finally,the activation threshold of object re-detector is determined,and the candidate region method combined with correlation filtering method is used to achieve re-detection,so as to achieve the purpose of long-term tracking.We conduct experiments using traditional and convolutional features with respect to the OTB2013 standard dataset and observe that the average success rates of tracking are 5.6% and 7% higher,respectively,when compared with that of SRDCF.Therefore,the proposed approach is a robust method for handling background blur,rotations,occlusions,and out-of-view objects.