基于金字塔特征的核相关滤波跟踪算法

Pyramid Feature Based Kernelized Correlation Filter Tracking Method

针对核相关滤波跟踪算法(Kernelized Correlation Filter tracking,KCF)对特征表达敏感的问题,提出了基于金字塔特征的KCF跟踪算法。该算法通过实验分析发现不同空间尺度上的HOG特征具有不同的目标-背景判别力和定位能力。基于该观测,融合多个尺度上的HOG特征提出了一种金字塔HOG特征,并将该金字塔特征应用于多通道的KCF跟踪框架。实验分析表明,该金字塔特征有效提高了KCF在多达50个视频序列上的跟踪性能。相比于原始KCF算法,所提算法在跟踪精度和成功率典型值上的性能提升分别为5%和4.6%。

Recently,kernelized correlation filters （KCF） have been successfully applied in vision tracking.However,more effective features are appealing to KCF trackers.In this paper,we proposed a pyramid feature based KCF tracking method （PKT）.The proposed method argued that the target-background discriminating and locating abilities of HOGs in different space scale were difference.A pyramid HOG feature （PHOG） was designed to achieve the both abilities.Then the proposed PHOG feature was applied to multi-channel KCF framework.Experimental results showed that the PHOG was effective for improving tracking performance which had been tested on 50 video sequences.Comparing with the original KCF,the proposed tracker had a significantly improvement of 5% in representative precision score and 4.6% representative success score.