基于核相关滤波器的多尺度目标跟踪

Multi-scale object tracking based on kernel correlation filter

目标尺度估计是视觉目标跟踪领域一个非常有挑战性的问题,现实场景中目标经常会发生不同的尺度变化,而目标的多尺度处理又极大增加了现有目标跟踪算法的耗时。近年来核相关滤波器因其高效性在目标跟踪领域引起了广泛关注,提出了一种实时的尺度自适应目标跟踪方法。首先由核相关滤波器对目标完成初步定位,再在尺度空间上训练一个独立的核相关滤波器,给出目标的精确尺度估计。此外,还提出了一种目标模型自适应更新机制,有效抑制了模型错误更新。实验表明,该方法在尺度变化、光照变化等各种目标外观变化情况下都有较好的跟踪性能,平均跟踪速度为101 f/s,可以满足实时应用的需求。

Object scale estimation is a challenging problem in visual object tracking,where in recent years kernel correlation filter is popular with its high efficiency. Thus,a real-time scale adaptive tracking method is proposed in this paper. Firstly,the target is initially located by the kernel correlation filter. Secondly,an independent kernel correlation filter is trained in the scale space to yield the precise scale estimation. Finally,a model adaptive update mechanism is proposed based on the scale estimation,which restrains the update error. The experimental results show that the proposed algorithm can perform superiorly under the condition of different appearance variations,such as scale variation,illumination variation etc. The average tracking speed can keep in 101 f/s,which meets real-time application requirement.