Deep neural network for person re-identification in a non-overlapping camera network

Abstract

Person re-identification is important and challenging parts in a non-overlapping camera network. In this paper, we propose the person re-identification framework which consists of kernel size into convolutional layers considering the person ratio and relationship matrix that train the relationship information related to neighborhoods. Our framework deals with global feature extracted from the whole body. The features generated by suitable kernel size are different to the local featured making by separated body images. The approaches of local feature extracted from divided bodies tend to lose salient information because of cutting the characteristic of products. The extracted features are used as elements to learn a relationship matrix which plays a role in distinction function. Our proposed framework outperforms state-of-the-art methods on challenging datasets.