一种改进的JSD距离的空间直方图相似度度量及目标跟踪

空间直方图是直方图的一种推广, 它能更精确地描述图像(或目标), 被应用到目标跟踪和图像检索等多个领域, 选择一种合适的度量两个空间直方图之间相似性的方法至关重要. 本文提出一种基于改进Jensen-Shannon divergence (JSD) 距离的空间直方图相似性度量, 将空间直方图中每个区间所对应像素的颜色特征和空间特征的联合分布看作一个带权重的高斯分布, 然后计算两个空间直方图对应区间之间的相似度, 即计算两个带权重的高斯分布之间的改进的JSD距离. 本文在计算JSD距离时充分利用高斯分布的权重, 从而提高了度量方法的区分能力. 理论和实验证明了本文提出的相似性度量的区分能力优于Ulges的度量方法, 视频跟踪结果也更稳定、更精确.

An Improved Jensen-Shannon Divergence Based SpatiogramSimilarity Measure for Object Tracking

Spatiogram is a generalization of histogram, which can more accurately describe the image (or the object) and has found application in object tracking and image retrieval. Suitable similarity measure is critical. In this paper, we present an improved Jensen-Shannon divergence (JSD) based spatiogram similarity measure. The spatial-color joint probability distribution of the pixels corresponding to each bin is regarded as a weighted Gaussian distribution. Then, the similarity of corresponding bin of two spatiograms is computed by improved Jensen-Shannon divergence with two weighted Gaussian distributions. We make full use of the weight of Gaussian distribution in order to improve the discrimination of the proposed measure. The proposed measure gives superior discriminative power both theoretically and experimentally than the measure proposed by Ulges, and the result of object tracking is more stable and accurate.