基于多特征信息和直方图相交的改进Meanshift算法

针对Mean-Shift算法仅利用颜色信息并以Bhattacharya系数为相似性度量导致算法准确率较低的问题,提出一种结合颜色和纹理信息并以直方图交集为相似性度量的跟踪算法。首先用局部二元模式算子提取纹理信息,再以对数比加权直方图取代传统直方图构建目标和候选目标模型,然后根据场景自适应地综合两类特征并通过Mean-Shift算法获得目标的初略位置,最后以直方图交集作为相似性度量并用Powell方法求其极值作为目标在当前帧中位置的估计。实验表明,该算法不仅能增强跟踪的准确性,而且对光照弱、目标与背景颜色近似等情况也有较好的鲁棒性。

Improved Mean Shift Algorithm with Multi-cue Integration and Histogram Intersection

The mean shift tracker is commonly used in real-time target tracking.However,the original mean shift tracker employs only color feature and uses the Bhattacharya coefficient as similarity measure,resulting in low tracking accuracy.This paper proposed a novel tracking algorithm,which integrated color and texture features and employed histogram intersection and Powell’s method to track.Firstly,texture feature was extracted by the Local Binary Pattern texture operator and integrated with color feature adaptively.Log-likelihood ratio histogram was proposed to represent objects instead of histogram.Then,the rough location of the target was obtained by the mean shift algorithm based on the two features.Finally,histogram intersection was defined as the similarity metric between the target model and candidates and iteratively maximized by Powell’s method.Experimental results demonstrate the proposed method can track targets more accurately and fast.