自适应尺度的快速相关滤波跟踪算法

Fast Scale-Adaptive Correlation Tracking

针对常规的相关滤波跟踪算法不能很好地适应目标尺度变化,而基于尺度金字塔的相关滤波跟踪算法虽然取得较高的跟踪精度,却在跟踪速度上大打折扣的问题,提出一种简单快速的自适应目标尺度的相关滤波跟踪算法.首先对目标模板进行对数极坐标变换,把目标的尺度变化转化为位移信号;然后对目标模板变换前后分别提取HOG特征,并建立位移与尺度的滤波模型;最后在相关滤波框架下同步跟踪目标的位移与尺度因子,并将二者融合得到目标跟踪框.实验结果表明,该算法的跟踪精度略低于基于尺度金字塔的相关跟踪算法,而跟踪速度却达到后者的2倍以上.

Since conventional correlation tracking algorithm cannot adapt to scale variations of the target well,and the scale-pyramid based correlation trackers gain higher precision but sacrifice more speed, this paper proposes a simple and efficient scale-adaptive correlation tracking method. Firstly, we converted the target scale to translation signals by taking log-polar transformation. Then, we extracted HOG features from target in Cartesian and log-polar coordinates, and built translation and scale correlation filter models. Finally, the translation and scale factor were tracked synchronously in the framework of correlation filter and fused into tracking bounding rectangle. Experiment results demonstrate that our tracker achieves much little loss in precision comparing with scale-pyramid based correlation trackers, but performs more than twice as the latter ones in speed.