Identifying a person across cameras in disjoint views at different time and location has important applications in visual surveillance. However, it is difficult to apply existing methods to the development of large-scale person identification systems in practice due to underlying limitations such as high model complexity and batch learning with the labeled training data. In this paper, we propose a prototype system design for large-scale person re-identification that consists of two phases. In order to provide scalability and response within an acceptable time, and handle unlabeled data, we employ an agglomerative hierarchical clustering with simple matching and compact deep neural network for feature extraction.
展开▼
