采用多相关滤波策略的鲁棒长时自适应目标跟踪  

Robust Long-term Adaptive Object Tracking Based on Multi-correlation Filtering Strategy

作　　者：谭建豪[1,2] 殷旺 刘力铭 王耀南 TAN Jian-hao;YIN Wang;LIU Li-ming;WANG Yao-nan(College of Electrical and Information Engineering,Hunan University,Changsha 410082,China;National Engineering Laboratory for Robot Visual Perception and Control Technology,Changsha 410082,China)

机构地区：[1]湖南大学电气与信息工程学院,长沙410082 [2]机器人视觉感知与控制技术国家工程实验室,长沙410082

出　　处：《计算机科学》2020年第12期169-176,共8页Computer Science

基　　金：国家自然科学基金(61433016)。

摘　　要：传统相关滤波方法在目标运动模糊和光照变化上取得了一定的鲁棒效果,但当目标存在形变、颜色变化、重度遮挡等干扰因素时难以实现跟踪,鲁棒性差,且当目标丢失后不能再恢复,无法实现长时间跟踪。因此,文中提出了一种鲁棒长时自适应目标跟踪算法。首先,提出了一种特征互补策略,将方向梯度直方图和全局颜色直方图的特征响应线性加权,学习对颜色变化和形变都具有鲁棒性的相关滤波模型,用以估计目标位移;然后,仅提取目标前景HOG特征,学习一个判别滤波器,用以保持对目标外观的长期记忆,使用该长期滤波器的输出响应来判别是否出现遮挡或跟踪失败,采用在线SVM分类器对丢失目标进行再检测,从而能够跟踪已丢失目标,以实现长期跟踪;其次,学习了以目标位置为中心的特征金字塔模型以预测尺度变化,防止目标框漂移;最后,在OTB目标跟踪基准数据集上对算法进行实验,并与目前较为流行的目标跟踪算法进行对比,进一步验证了所提算法的鲁棒性、准确性和优越性。The traditional correlation filtering methods have recently achieved excellent performance and shown great robustness to exhibiting motion blur and illumination changes.However,it is difficult to achieve tracking when the object has interference factors such as deformation,color change,and heavy occlusion.It shows poor robustness when the object is lost and cannot be recovered to achieve long-term tracking.Therefor,this paper proposes a robust long-term object tracking algorithm.First,a feature complementation strategy is proposed,which linearly weights the feature responses of the directional gradient histogram and the global color histogram,and learns a correlation filtering model that is robust to color changes and deformations to estimate the target displacement.Then,the object features are taken to learn a discriminant correlation filter to maintain long-term memory of object appearance.We use the output responses of this model to determine if tracking failure occurs.We use the online SVM classifier to re-detect the lost objectand retrack the lost target which can effectively recover the tracking target from failure to achieve long-term tracking.In addition,this paper learns a correlation filter over a feature pyramid centered at the estimated object position for predicting scale changes and further enhance robustness and accuracy.Finally,this paper compares the proposed algorithm with the state-of-the-art performance tracking algorithms on the online object tracking benchmark.The result shows that the proposed algorithm performs great robustness and accuracy.

关 键 词：长时目标跟踪 颜色直方图 相关滤波 SVM再检测器 尺度自适应 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]