基于线性组合预测和生物特征的人体肤色检测

Human Skin Detection Based on Combination of Forecasts and Skin Biological Features

针对视频监控中基于肤色特征人体目标检测中的两个棘手问题,即人体肤色受光照变化影响较大,以及复杂背景下肤色相近色的干扰,提出了一种新的肤色检测方法。首先假设视频序列每帧肤色区域像素在彩色空间的分布构成相对集中的"点云"三维几何体,光照变化时每帧"点云"几何体在彩色空间的变化可以通过平移、缩放和旋转等参数约束下的三维仿射变换来建模,提出了用线性组合预测模型来预测这三类参数的变化,进而预测并更新待检测帧的直方图分布;然后利用Bayes分类器进行肤色区域的初分割。为了克服复杂背景中肤色相近色的干扰,本文采用组合彩色空间变换凸显人体肤色生物特征,减少了肤色和非肤色在单个彩色空间时的重叠区域,在初分割的基础上进一步消除大片相近色的干扰。经过大量实验证明,该方法在帧间光照变化的情况下对肤色变化有很好的敏感性,且能有效克服大片背景相近色的干扰。

Illuminanee variation and camouflage disturbance from the background are two intractable problems for human skin color detection in surveillance, a new method is proposed for detecting this problem. Firstly, it is hypothesized that skin pixels in each frame are compact together as a ＂dot cloud＂ in a color space ; the shape evolution of each frame of the ＂dot cloud＂ is parameterized as the translation, the scaling and the rotation; the combination of forecasts method is proposed to predict these arguments for the next frame being segmented, thus its histogram can be figured out. Then, Bayes classifier is used to obtain the primary skin patches. Finally. human skin biological features and the combination of color spaces are adopted to eliminate the interruption of the camouflage. Extensive tests prove its sensitivity to the human skin color, and it is robust to the mass camouflage perturbation from the background.