基于全局空间约束块匹配的目标人体识别

Patch Matching with Global Spatial Constraints for Person Re-Identification

目标人体识别即非重叠多摄像系统中人的重现(person re-identification)问题,当前多数的目标人体识别都是通过提取人体表观特征,并利用特征的相似性对目标人体进行重识别,这些方法对于一些大部分表观区域相似而小部分区域不同的行人仍然无法给出准确的识别结果.考虑到目标人体识别中的行人几乎都处于站立姿势,同一行人的不同图像在垂直方向上的全局结构比不同行人间的更加相似.在基于稠密块匹配的基础上,提出了全局空间约束块匹配的识别方法.该方法不仅考虑2幅图像中局部块的匹配,还考虑各块在自身图像中垂直方向上的全局空间约束.为了减少背景对识别的负面影响,采用姿势评估的方法来提取大致的人体前景.在实验中,提出的方法在经过最具挑战的公用VIPeR数据库和CUHK02数据库测试后,该方法对人体识别率起到了显著的改善作用.

The target person recognition is a problem of person re-identification in multiple nonoverlapping camera views.Existing target person recognition mostly extracts the human appearance feature,and re-identify target pedestrians through the feature similarity.For the pedestrians who have the most similar area and a small different part,these methods still can not give accurate recognition results.In this article,we consider that pedestrians of recognition are almost in a standing posture,and the vertical structure of the same pedestrian is more similar to the vertical structure of different pedestrians.Therefore,on the basis of densely patch-matching,we propose a matching method with spatial constraints（SCM）,which not only considers the process of local patch matching in two different images,but also concerns the constraint of each patch in the vertical direction.In order to reduce the negative impact of background for identification,we adopt the method of pose evaluation to extract roughly foreground of the human body.In our experiment,the proposed approach have been tested in the most challenging public VIPeR database and CUHK02 database,and the results prove that it reaches the best recognition results so far.