• face detection
• face recognition
• LBPH
• OpenCV
• PCA
• Smart camera networks

The aim of this work is to design and develop a face recognition system running on smart camera
networks. In many systems, these are passively used to send video to a recording server. The
processing of the acquired data is mainly executed on remote and more powerful computers (or
clusters of computers). In this thesis a distributed architecture was developed where computer
vision algorithms are executed on smart cameras, which can exchange information to improve
resource balance. A smart camera network has been defined specifying the roles that client
nodes and a server have, and how nodes cooperate and communicate among them and with the
server. Smart cameras, initially look for changes in the environment. When motion is detected,
they perform face detection. Once a face is found, the camera itself processes it and tries to
asses whom it belongs to, using a local cache of recognizers. This cache stores a portion of the
whole information present on server side, and can be used to perform recognition tasks on the
smart cameras. If a node is not able to identify a face it sends a query to the server. Finally,
if the person’s id can be determined, either by the server or the client itself, the occurrence of
the correspondent recognizer is notified to the nearest nodes. Human faces that were not recognized,
are stored on the remote server and can be manually annotated. Clustering algorithms
have been tested in order to automatically group faces belonging to unknown people on server
side so they made the manual annotation easier. Extensive experiments have been performed
on a freely available dataset to both assess the recognition performance and the benefits of
using collaboration among cameras. Raspberry PI devices were used as camera network nodes.
Various tests were performed in order to verify the efficiency of the face recognition approach
on such devices.