融合底层和中层字典特征的行人重识别

针对当前行人重识别方法采用单一底层特征识别率较低的问题,提出一种融合底层和中层特征的识别方法,由粗到精对人体目标进行匹配识别。首先提取目标的颜色直方图和纹理直方图进行粗分类;然后将人体目标分为头部、躯干和腿部3个部分。忽略包含信息量较少的头部,对躯干和腿部,提出一种中层图像块字典提取方法,并对照该字典生成中层特征,进行精确分类。底层特征结合中层特征使算法既具有较好的区分度,又具有良好的泛化能力。实验结果表明本文算法在VIPeR数据库上的n AUC比已有方法提高6.3%,对遮挡和背景粘连的鲁棒性更好。

Pedestrian re-identification based on fusing low-level and mid-level features

Aiming at the problem of low recognition rate in the existing pedestrian re-identification algorithm using single low-level feature, a new method by fusing low-level and mid-level features is proposed, which identifies person in a coarse to fine strategy. First, the pedestrian is recognized roughly by color and texture features. Then, the human body is divided into three parts, including head, main body and leg. Head is ignored for its few useful information. A mid-level dictionary method is proposed and the dictionary is trained using patches from main body and leg, and then mid-level feature is computed for fine recognition. Fusing mid-level and low-level features can be not only discriminative but also representative. The experimental results indicate that the proposed method can increase nAUC by 6.3% compared with the existing methods, which is more robust to occlusion and background adhesion.