引入垂直全局表观约束的中层特征行人再识别.pdf

研究与开发 引入垂直全局表观约束的中层特征行人再识别 王亚东，张荣，蒋桧慧，郭立君 （宁波大学信息科学与工程学院，浙江 宁波 315211 ） 摘 要：行人再识别问题中，包含语义信息的中层特征能够提供更强的判别力。由于中层特征也采用局部匹 配方式，与底层特征一样存在由于不同行人部分表观区域比较相似而产生误匹配问题。考虑到行人几乎都处 于站立姿态，同一行人在垂直方向上的表观序列比不同行人的更相似，提出了在中层特征的基础上引入行人 垂直全局表观约束，并融合底层稠密块匹配的识别方法。实验结果表明，算法在最具挑战的公用 VIPeR 数据 库和 CUHK01 数据库上，均取得了比现有方法更高的命中率。 关键词：行人再识别；中层特征；垂直表观约束 中图分类号：TP391 文献标识码：A doi: 10.11959/j.issn.1000−0801.2017023 Person re-identification using mid-level features with vertical global appearance constraint WANG Yadong, ZHANG Rong, JIANG Huihui, GUO Lijun College of Information Science and Engineering, Ningbo University, Ningbo 315211, China Abstract: Mid-level features with the semantic information can provide stronger discrimination in the person re-identification than low-level features. But like the low-level features, the mid-level features also use local feature matching methods and easily leads mismatch problem when different pedestrians have similar appearance features in some local areas. Considering the same pedestrians are more similar than different pedestrians in the vertical direction since pedestrians are almost always in a standing position, global vertical appearance constraints was introduced. Furtherly, a method for person re-identification was proposed which fuses the low-level densely patch-matching and the mid-level features with the global vertical appearance constraints. Experimental results show that the proposed algorithm can achieve a higher hit rate than the existing methods on the most challenging public VIPeR database and CUHK01 database. Key words: person re-identification, mid-level feature, apparent constraint 收稿