A Study of Person Re-Identification System

Person Re-Identification (Re-ID) consists in recognizing an individual who has already been observed over a network of non-overlapping camera views (hence the term Re-Identification). However, the matching is challenging due to similarities of person’s appearance across different cameras. Recent advances have shown that metric learning methods are effective for person Re-ID as they provide a robust metric for measuring (dis)similarities among (un)matched image pairs. However, these methods suffer from the Small Sample Size (SSS) problem due to the limited number of labelled training samples. This paper provides an overview of the classical approaches in Re-ID which consist in exploiting the appearance cues such as color or texture of clothing. Then, we navigate to the metric learning-based methods which consist to establish the corresponding/matching using matching function (similarity metric or a ranking function) of appearance signatures. In this direction, we start our discussion with appearance-based methods based on hand-crafted feature and the metric learning-based approaches. The relevant Re-ID approaches are described in detail. We present also the concept of RGB-D based Re-ID and we summarize the most recently work utilizing color-depth sensors. The commonly used benchmark datasets for person Re-Identification are summarized and discussed. The performance of some state-of-the-art person Re-ID approaches on most used benchmark datasets is also compared and analyzed.