| Description: |
Distributed communication networks offer the advantages of scalability and not having a single point of failure. In distributed camera networks, each camera is equipped with an on-board processor capable of analyzing its images locally thus allowing most communication to take place between nearby cameras, with only summarized information being transmitted to a central server. Based on the wireless sensor cluster-based protocol developed by Medeiros et al. (2008), this thesis presents a multi-person tracking framework that represents individual observations of face pose from multiple cameras in a shared world coordinate system, allowing observations to be compared and integrated without reference to individual camera parameters. We show that the cluster-based protocol is useful for multi-object tracking not only in wireless camera networks, but also in wired camera networks. From a theoretical perspective, we show how the cluster-based protocol addresses computer vision challenges in wired networks using the same mechanisms that address communication challenges in wireless networks. From an empirical perspective, we demonstrate distributed tracking of people in real time in a twelve-camera wired network. This dissertation also proposes a benchmark for unconstrained face recognition based on human familiar face recognition. Face recognition holds the potential for greatly improving person tracking. However, current face recognition algorithms do not work reliably in the unconstrained images captured by camera networks. Our benchmark would allow future face recognition algorithms to be tested on the sorts of challenging images captured in a camera network and, at the same time, allow a comparison of the performance of these algorithms to unconstrained face recognition by humans. |