激光与光电子学进展, 2019, 56 (20): 201501, 网络出版: 2019-10-22
姿态引导与多粒度特征融合的行人再识别 下载: 1052次
Posture-Guided and Multi-Granularity Feature Fusion for Person Reidentification
摘要
行人再识别系统中,检索到的行人图像会出现较大的姿态差异、复杂的视角变化以及检测框中行人图像不对齐等问题,为此,提出一种可以直接使用人体关键点信息进行行人图像对齐,并在此对齐基础上提取多粒度特征的重识别算法。首先使用姿态预估模型定位人体骨架关键点信息,并根据提取的骨架关键点直接进行行人图像对齐,然后对行人图像提取多粒度特征。评估阶段使用姿态信息结合多粒度特征进行相似度匹配。仅使用身份(ID)损失函数在Market1501、CUHK03、DukeMTMC-reID 3个公开数据集上进行实验。结果表明,所提算法具有一定的优势。
Abstract
In the person reidentification system, the retrieved person image will have large posture differences, complex changes in perspectives, and misalignment of person images in the detection frame. In order to solve these problems, a reidentification algorithm is proposed,which can directly use the key point information of the human body for person image alignment and extract multi-granularity features based on this alignment. First, the posture prediction model is used to locate the key points of the human skeleton, and the person image is directly aligned according to the extracted skeleton key points, and then the multi-granularity features are extracted from the person image. The evaluation phase uses posture information combined with multi-granularity features for similarity matching. The experiment is carried out only using the identity(ID) loss function on the three public datasets of Market1501, CUHK03, and DukeMTMC-reID. The results show that the proposed algorithm has certain advantages.
张良, 车进. 姿态引导与多粒度特征融合的行人再识别[J]. 激光与光电子学进展, 2019, 56(20): 201501. Liang Zhang, Jin Che. Posture-Guided and Multi-Granularity Feature Fusion for Person Reidentification[J]. Laser & Optoelectronics Progress, 2019, 56(20): 201501.