Person re-identification with activity prediction based on hierarchical spatial-temporal model

Abstract Person re-identification (re-id) across cameras remains a very challenging problem, especially when the wide range searching exists in a multi-camera surveillance network. Current person re-identification methods focus on using visual model to search the specified person. In fact, in practical applications, due to the large-scale search range, the searching way only relying on visual model is not efficient. Moreover, the recall ability of visual model usually is limited in large-scale searching, because it does not consider the spatial-temporal information of person. However, the current public re-id datasets only include the visual samples. To address this problem, in this work, we collect a large-scale re-id dataset, PKU-SVD-B-REID, which includes both visual and spatial-temporal information of over 133 K samples. Then, we propose a novel person re-id framework, named Hierarchical Spatial-Temporal Model (HSTM), which can effectively predict the person activity path and reduce the search range in the real multiple cameras surveillance system. Extensive experiments on PKU-SVD-B-REID validate the superiority of our method over conventional re-id methods based on only visual information in terms of both efficiency and accuracy.