多特征融合的尺度自适应KCF人脸跟踪

A scale adaptive KCF human face tracking algorithm based on multi-feature fusion

针对传统核相关滤波(KCF)跟踪算法在人脸跟踪中无法处理尺度变化、严重遮挡等问题,提出一种多特征融合的尺度自适应KCF人脸跟踪算法。该算法先对肤色与HOG特征进行融合来表征人脸,通过多通道相关滤波器定位人脸位置;学习一个一维的尺度滤波器来估计人脸的最优尺度;采用线性插值的方式对滤波器系数和人脸外观模型进行更新。实验结果表明,改进后的算法能明显提高跟踪的性能。通过定量与定性分析该算法对尺度变化、严重遮挡等问题有很好的鲁棒性,跟踪速度在36.7 f/s时达到实时应用的要求,优于近几年一些优秀的跟踪算法。

Since the traditional kernel correlation filtering(KCF)tracking algorithm cannot handle the problems of scale variation and heavy occlusion during human face tracking,a scale adaptive KCF face tracking algorithm based on multi-feature fusion is proposed.In the algorithm,the human face is represented by fusing the skin color with HOG features.The human face position is located by using the multi-channel correlation filter.The optimal scale of human face is estimated by learning a one- dimensional scale filter.The linear interpolation mode is adopted to update the filter coefficient and facial appearance model. The experimental results show that the improved algorithm can significantly improve tracking performance,has good robustness for problems of scale variation and heavy occlusion by means of quantitative and qualitative analysis,and can meet the real-time application requirement at the tracking speed of 36.7 f/s,which is superior to some excellent tracking algorithms in recent years.