Master thesis : Person re-identification in a camera network

Abstract

In the modern world, data collection is increasingly used to gain insights into our lifestyles. Companies are tracking us in every way imaginable to fully understand our needs. Smart cameras are one method that can be used to capture new information about customer behavior. In this thesis, we provide an implemented approach to detect, track, and re-identify people using live streams from non-overlapping surveillance cameras. To do so, we combine the well-known computer vision tasks of detection and tracking. In addition, we include a re-identification principle that allows a person to be re-identified as they move from one camera to another. In other words, we allow the assignment of a unique identifier to each tracked person and the retrieval of a person’s identifier when they have not been observed by a camera for a certain period of time. In our case, person re-identification relies heavily on a fairly recent deep learning model, the OSNet-AIN model. It provides essential information about the appearance of different people. These data obtained allow us to differentiate one person from another through the different cameras. However, other information such as the movement of people is also taken into account to improve this re-identification. Evaluated on videos provided by ARHS Spkiseed, the implemented method realizes only one error during re-identification. We also obtain frame rates between two and six taking into account that no GPU was used. The algorithm can therefore work in real time.