People Re-identification across Non-overlapping Cameras using Group Features

Norimichi Ukita   Yusuke Moriguchi   Norihiro Hagita

Abstract

This paper proposes methods for people re-identification across non-overlapping cameras. We improve the robustness of re-identification by using additional group features acquired from the groups of people detected by each camera. People are grouped by discriminatively classifying the spatio-temporal features of their trajectories into those of grouped people and non-grouped people. Thereafter, three group features are obtained in each group and utilized with other general features of each person (e.g., color histogram, transit time between cameras, etc.) for people re-identification. Our experimental results have demonstrated improvements in people grouping and people re-identification when our proposed methods have been applied to a public dataset.

Manuscript

Citation