采用核相关滤波器的自适应尺度目标跟踪

Adaptive scale object tracking with kernelized correlation filters

由于现存的大多数基于检测的跟踪器都没有解决尺度变化问题,本文在传统的基于检测的目标跟踪框架下设计了一种尺度估计策略,并给出了基于核相关滤波器的自适应尺度目标跟踪算法。该算法利用核函数对正则化最小二乘分类器求解获得核相关滤波器,通过对核相关滤波器的在线学习完成目标位置和尺度的检测,并在线更新核相关滤波器。为了验证本文算法的有效性,选取了10组场景复杂的视频序列进行测试,并与其它5种优秀跟踪方法进行了对比。结果表明,本文提出的方法比上述5种优秀跟踪方法中的最优者的平均距离精度提高了6.9%,且在目标发生尺度变化、光照变化、部分遮挡、姿态变化、旋转、快速运动等复杂场景下有较强的鲁棒性。

As most of tracking-by-detection methods have not dealt with the scale estimation problem in target tracking, this paper proposes a scale estimation strategy based on the tracking-by-detection framework. Meanwhile, it designs an adaptive scale tracking algorithm based on kernelized correlation filters. The algorithm uses a kernel function to solve the regularized least square classifier to obtain the kernelized correlation filters. Then it completes the target position and scale detection by online learning the kernelized correlation filters, and updates the filters in online. To verify the feasibility of the proposed algortihm, ten groups of benchmark video sequences are tested and obtained results are compared with those of five kinds of tracking algorithms. The experimental results show that the proposed approach improves the performance by 6.9% in the average distance precision as compared to the best one of the other five excellent existing tracking algorithms. It is robust to scale changing, illumination variation, partial occlusion, pose variation, rotation, fast motion and other complexscenes.