Detecting humans in a cluttered scene
Jérôme Vignola
Robert Bergevin (Supervisor)
Problem: Tracking people is a very active problem in computer vision. The goal of this project is to create a system to robustly detect humans in a cluttered scene and create a 2D skeleton of the detected persons. Detection is made on the basis of a single colour image.
Motivation: This project is part of a project called COGNOIS, which aims at localizing humans in a cluttered scene to eventually track them and understand their actions. This project is one of the first steps in COGNOIS. Information gathered during this research will be used, in a further step, to create a 3D representation of skeleton type using stereoscopy.
Approach: Approach used in the project is general, in the sense that no hypothesis on identity or appearance of the person is made a priori. Detection is made using a simplified model of human morphology and possibly colour just as well as some heuristics. Basic knowledge is given to the system, such as perceptual grouping laws, and the system must establish links between the elements to perform human detection and recognition of their parts. All of this is done without any specific model, to make the system as generic as possible.
Challenges: Many human detection systems already exist. However, they often have many constraints and strong hypotheses, such as wearing tight clothes, walking parallel to the camera, staying upright, having a fixed background, etc. In addition, the majority of these systems track people without localizing their parts, which is insufficient for our needs. Our biggest challenge will be to create a robust system without a priori hypotheses on the person's clothes or position, etc. This is to ensure that the envisioned real world applications world are feasible.
Applications: Applications of such project are numerous. For instance, parking, airports and other public areas surveillance, perceptual man-computer interfaces, learning of sport movements, assisted car driving, etc.
Calendar: September 2001 to December 2003
Support: This research is supported by Robert Bergevin (NSERC individual research grant) via the COGNOIS project.
Last modification: Sep 28 2007 2:09PM by vignolaj


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