| Region matching of people images taken at different times and locations |
| Masters |
| Michel Lantagne |
Robert Bergevin (Supervisor)
Marc Parizeau (Co-supervisor) |
|
| Problem: With the recent availability of low cost yet powerful computer hardware, one can now envision the emergence of sophisticated and intelligent surveillance systems integrating a network of loosely-coupled computation nodes, each connected to a camera. These systems would need to track a person from non overlapping fields of view in order to determine whether each camera is observing the same person. The problem consists of measuring the similarity between two human silhouettes. |
| Motivation: This project is a part of COGNOIS: Communication and Observation toward a Generic Natural Ontogeny for Intelligent Systems. The goal of COGNOIS is the development of a general intelligent architecture with skills for observation and communication for many applications in the real world. One application of COGNOIS is the construction of a system able to detect and track one or many persons in a scene with a sparse network of cameras. Thus, this system requires a module for information integration of images taken at different times and viewpoints. |
| Approach: Human silhouette comparison can be addressed by characterizing a person’s appearance. The project takes in entries of a human silhouette segmented image which may include labelled body parts. The approach uses different methods of region matching. The project is divided into several parts. First, many descriptors of colour and texture are studied and tested so as to enable the characterization and segmentation of human body parts. Second, a region matching scheme is used to compare the regions within two human silhouettes. Then, a similarity measure is defined to facilitate person matching. Finally, spatio-temporal coherence is used to improve the matching and tracking. |
| Challenges: Many systems have been developed recently for the detection and tracking of people, but in the majority of cases they impose many constraints. For example, several systems assume that there is only one person in the scene at a time, the person must stay within the scene and face the camera, the background must remain static or simple, etc. When these assumptions are not respected, the system’s robustness degrades rapidly. The main challenge is to develop a system where the number of starting assumptions is reduced while preserving a good robustness. |
| Applications: Description and tracking of people is useful in several situations. Here are some examples: intelligent monitoring and surveillance systems in airports, parking lots and old age homes for detection of problematic situations (and to start an alarm), user interfaces allowing actions to be defined by simple movements of the user, participation and interaction in a virtual reality, translation of sign language, supervision and interventions in medical operations, etc. |
| |
| Calendar: September 2001 - September 2003 |
| |
| |
| Last modification: 2007/09/28 by lantagne |
