基于多图学习的无约束行人再识别

Person re-identification is the task of matching a probe subject with the subjects in the gallery from different cameras. As one of the most important techniques in security and surveillance, person re-identification has received a lot of attentions in both academia and industry. Existing methods mainly focus on feature design or distance metric learning for improved matching. In practice, due to the variations in resolution, pose, illumination, scale, and occlusion in the captured images by different cameras, accurate person re-identification is a fundamentally challenging task, and existing methods are insufficient to achieve desired accuracy. In light of this, we propose an unconstrained person re-identification method based on multi-graph learning. Specifically, to achieve robust and accurate matching, we utilize multiple feature representations extracted at different scales for accurate local matching. Then, the global matching is achieved by learning the similarities between probe and gallery through multi-graph optimization and fusion.

行人再识别是指给定查询行人图像，在已有的行人数据库中识别出与此人匹配的图像。作为安全监控领域的一项重要技术，行人再识别在学术界和工业界都得到了广泛关注。现有的方法多是基于特征描述设计或者距离度量学习来实现更准确的匹配。但在实际应用中，由于不同相机中所获取的图像存在分辨率、行人姿态、光照、尺度、遮挡等差异，精准的识别并匹配不同相机中的行人是一项非常困难的任务，而现有方法难以达到较高精度。针对这一问题，本课题提出在多图学习框架下的无约束行人再识别方法，利用在多尺度下提取的多样特征，首先实现查询图像和行人数据库中图像局部特征的精确匹配，然后通过多图优化和融合的手段来学习查询图像和行人数据库中图像的全局相似度，从而实现鲁棒精准的行人再识别。

随着热点地区人口密度的加大以及视频监控设备的大范围部署，出于社会安全的考虑，人群监控和视频分析在近些年有了极大的需求。其中一项重要的任务就是行人再识别。该任务是指给定查询行人图像，在已有的行人数据库中识别出与此人匹配的图像。早期提出行人再识别的方法首先基于行人检测，并在检测到的区域进行传统视觉特征提取，包括颜色，材质等特征，而匹配则是基于对行人特征的度量计算。在实际应用中，受到图像分辨率、姿态、光照、尺度、遮挡等影响，这些方法普遍难以达到精准的效果。 针对这一问题，本课题提出在多图学习框架下的无约束行人再识别方法，对于行人图像进行多尺度特征提取，然后将行人匹配这一问题转化为为图优化问题，并融合多种特征，构建多图同时优化，从而实现鲁棒精准的的行人再识别。该项目实验基于学术界公开数据集，识别精度达到了领先水平，成果发表在IEEE Transactions on Circuits and Systems for Video Technology，IEEE Transactions on Neural Networks and Learning Systems等一流期刊以及学术会议。于此同时，研究成功具有实际应用价值，能够部署在现有视频监控系统中。