Tracklet style transfer and part-level feature description for person reidentification in a camera network

摘要

Reidentifying multiple objects in a camera network are a difficult problem, especially when determining whether the same object appears in a different place at a different time, captured by another camera. In this paper, we propose a novel tracklet-based approach for reidentifying objects despite the illumination condition differences that occur at various times of day. A similarity search is performed by comparing part-level object feature descriptions. Tracking in each camera is made by a recurrent neural network, and the matching between cameras is done by using a similarity neural network to obtain an output in the form of a similarity score. Our approach consists of two main phases. In the first phase, preprocessing is performed through the transfer of tracklets from several cameras. This process generates more samples from each camera, which is beneficial for training. In the second phase, the object definition is applied, which considers appearance information, temporal information and the similarity calculation, hence making object reidentification easier. We have analyzed the proposed strategy when applied to pedestrian reidentification databases in comparison with state-of-the-art work to prove its robustness.