16 Feb 2015 |
Research article |
Intelligent and Autonomous Systems
Video Surveillance Research at ETS LIVIA Laboratory
Header image bought from Istock website: Copyright protected.
Activities at the Laboratory for Imagery, Vision and Artificial Intelligence (LIVIA) of École de technologie supérieure (ÉTS) in Montreal focus on forms visual recognition and computer vision. In this context, six priority areas have seen in-depth research carried out over the years: i) artificial vision; ii) automatic image processing for documents; iii) imagery (medical, aerial, etc.); iv) biometrics, surveillance and intruder detection; v) learning in static and dynamic environments; and vi) perception and environment.
On February 18, 2015, Professor Éric Granger will give an overview of his current research on video surveillance face recognition as part of the Graduate Studies Evening Events (“Soirées des cycles supérieurs” in French). At this conference, he will present current research on various systems for adaptive face recognition video surveillance in changing environments. These systems can, for example, be used by public safety agencies in sensitive public places such as train stations and airports.
In this article, we will look at multi-classifier systems devised to identify individuals or follow their actions, even when the systems have very little data to work with. A subsequent article will look at adapting facial models and contextual recognition used in video surveillance.
Multi-classifier systems for video surveillance
A system for recognizing faces in video surveillance can detect the presence of persons of interest, who are part of one or several databases. A classifier is a specialized device that compares the discriminant representation of each image captured in video sequences with that of a reference image captured during the subscription. The term “multiple classifiers” means that these systems use multiple diverse classifiers operating in parallel in order to make the best decision.
Several classifier types exist. Here is three examples of multiple classifiers systems:
- A first type of multiple classifiers system divides the face into sub-blocks and dedicates both a classification technique and a classifier for each sub-block to make the system more robust to variations, because the videos follow moving individuals;
- A second type of multiple classifiers system works using images captured in movement. Specialized classifiers are created for different conditions (face possible angles of capture, brightness, etc.). Each classifier works in parallel and takes a decision on recognition;
- A third type of classifier uses algorithms to extract different types of information of a face (the characteristics of the skin, facial structure, etc.). It creates classifiers for each category of information that focus afterwards on their field of specialization in order to reach a joint decision on recognition.
All these classifiers act somehow as experts in a particular field who give their opinion on their own specialty. If the multi-classifier system discovers a new capture condition (difference in viewing angle, brightness, or another aspect not shown by the reference faces or by classifiers in place), one or several classifiers may be added to fill this gap.
The images shown below demonstrate how a classifier dedicated per person. The system starts by enrolling the person by entering an image of their face in the database.
In recognition mode, the system follows this person’s face as it moves and the prediction of the classifiers is accumulated according to each different person to produce robust results. The outline of the tracking rectangle turns red when the system determines that a face has been recognized as one of the individuals enrolled in the database.
Multi-classifier systems are useful tools for surveillance officers. They issue a recognition notification when there are sufficient indicators that a person within the database has been identified. An agent monitoring several screens will see that the system has recognized someone on one of the screens, and then the agent can determine whether or not to approach the person to confirm their identity.
Professor Granger’s Conference
On February 18, 2015, Professor Éric Granger will discuss LIVIA’s research into certain aspects of video surveillance as part of the Graduate Studies Evening Events (“Soirées des cycles supérieurs” in French. Admission is free. Click here to register.
Christophe Pagano is a PhD student in the Automated Manufacturing Engineering Departement at ÉTS. He is specialized in spatiotemporal systems for face recognition in video sequences.
Program : Automated Manufacturing Engineering
Research laboratories : LIVIA – Imaging, Vision and Artificial Intelligence Laboratory
Eric Granger is a professor in the Systems Engineering Department at ÉTS. His research focuses on machine learning, pattern recognition, computer vision, information fusion, and adaptive and intelligent systems.
Program : Automated Manufacturing Engineering