Reranking optimization for person re-identification under temporal-spatial information and common network consistency constraints

摘要

Recent research of person Re-identification (ReID) most focuses on exploring person appearance feature and distance measure between specific camera pair, and seldom considers complex camera network consistence and physical information as supplemental factors for improving the re-identification performance. In this paper, a re-ranking optimization framework under temporal-spatial information and common network consistency constraints is proposed to compensate the accuracy deterioration caused by appearance only based ReID methods in all pairwise cameras. Firstly, a correction function is introduced for describing the influence factor of temporal-spatial information on similarity scores. Then the amended similarity score strategy is provided to tradeoff between the person appearance and temporal-spatial information. Finally the global optimization problem of the jointing temporal-spatial and common consistence constraints is solved by integer programming algorithm. Furthermore, we try to solve the generalized situation where cameras and persons are more random by introducing topology information according to geometry around the cameras, taking the place of network consistency. The experiment results validate that the proposed framework significantly improves performance compared to the other person re-identification methods on the multi-cameras RAiD dataset and TMin dataset both in simple and generalized camera network. (C) 2018 Elsevier B.V. All rights reserved.