多特征融合的人体目标再识别

Person re-identification based on multi-features

在非重叠的多摄像机监控系统中,人体目标再识别是需要解决的主要问题之一。针对当前人体目标再识别使用目标的外观统计特征或者通过训练获取目标特征时存在的问题,提出一种无需训练,对视角、光照变化和姿态变化具有较强鲁棒性的基于多特征的人体目标再识别算法。首先根据空间直方图建立目标整体外观表现模型对目标进行粗识别,之后将人体目标分为3部分,忽略头部信息,分别提取躯干和腿部的主色区域的局部颜色和形状特征,并通过EMD(earth movers distance)距离进行目标精识别。实验结果表明,本文算法具有较高的识别率,且不受遮挡和背景粘连的影响。

In non-overlapping multi-camera surveillance systems person re-identification is one of the main issues. Aiming for person re-identification useing statistical properties of the objects and features by training, we propose a method by com-bining global and local features to identify the same person in different images. This method does not need a training phase, and it is robust to different viewpoints, illumination changes, and varying poses. First, the object is recognized roughly by spatiograms. Then the human target is divided into three parts. By ignoring the head part, the local color and shape fea- tures of the main body, the arms and the legs are extracted. Thus, the recognition of the person is carried out according to the Earth movers distance of the local features. The experimental results show that the proposed method has a higher accu- racy rate, and it is invariant to the effects of occlusion and background adhesion.