Incremental description of deformable objects in video sequences using computer vision
Stéphane Drouin
Marc Parizeau (Supervisor)
Patrick Hébert (Co-supervisor)
Problem: The objective of this research is to automatically describe the movements of an articulated object from a video sequence of its movements. This project contributes to establishing an autonomous system by proposing an automatic method of acquisition and the use of new knowledge.
Motivation: Intelligent systems are becoming omnipresent in our environment. Their interactions with people require them to have a capacity to acquire new knowledge through a learning process. The COGNOIS project was established in this context. One of the objectives of this project is the observation of people and the understanding of their behaviour by the intelligent systems.
Approach: The rigid parts of an articulated object are segmented on the basis of their movements, and these parts are automatically arranged in an articulated model which is then used to track the object. The automatic construction of a model has evolved due to the initialization problem encountered in all tracking systems and provides an aptitude to acquire new knowledge. The proposed approach is general and allows the description of several types of articulated objects, such as people or manufactured objects.
Challenges: The automatic tracking of an articulated object by a computer vision system must meet two objectives: the system must remain attached to the target from the beginning to the end of the sequence (tracking) and the system must estimate the temporal parameters which describes the movement (movement measurement). Achieving these objectives entails addressing the following three challenges: initialization, robustness and stabilization. Initialization includes several aspects which are necessary for the system to obtain an initial interpretation of the scene, such as camera calibration and modeling of the background. In existing tracking systems, the most problematic part involves the initialization of the model. Typically, this aspect is accomplished manually and includes: the choice of model to represent the data, the estimation of invariant parameters of the model (size and appearance of the limbs) and the initial correspondence between the model and the images. Robustness enables the tracking to continue despite the presence of undesirable elements in the images. In this way a robust system can resist to partial occultations, prolonged or not, and to resist total occultations by reinitializing the tracking at the end of the event. A robust system must also tolerate noise such as problems related to lighting, shadows or sensor noise. Finally, stabilization is a filter which makes the description coherent with the model used despite the presence of noise in the measurements.
Applications: One application of such a system is the surveillance of public areas. Within the COGNOIS project, a surveillance system exploiting sensorial, optic, acoustic, mechanical, etc. modalities, would be able to identify in real time any unexpected events and contact competent authorities when needed. For example, this system could detect a person who has fainted in a parking lot and request medical assistance.
Expected results: The main contribution of this research is the fact that it offers a choice of segmentation and tracking methods and can propose a general tracking system of articulated objects.
Calendar: January 2003 - December 2007
Support: NSERC, FQRNT, Precarn
Web reference:
Last modification: 2007/11/12 by sdrouin


©2002-. Computer Vision and Systems Laboratory. All rights reserved