Intelligent fusion of a hybrid (infrared and visible) sensor in the context of pedestrian detection and surveillance.
Hélène Torresan
Xavier Maldague (Supervisor)
Patrick Hébert (Co-supervisor)
Problem: The detection of the movement of people has become more and more important over the past few years. Numerous applications in the area of security and surveillance are emerging. The goal of this project is to develop a prototype combining an infrared and visible sensor to enable the detection and surveillance of pedestrians over a period of time. More specifically, the project will be carried out in an environment where one to three pedestrians are moving in a range of 9 to 21 meters within an area affected by various lighting and atmospheric conditions involving wind, snow, night and day.
Motivation: The addition of an infrared sensor will provide information which complements the images obtained in the visible range. Visible images offer a rich content where the detection of people can however be limited by a change in lighting conditions. Infrared images generally allow a better contrast to be obtained between a person and the environment, but these images are not as robust to changes in temperature and wind conditions. An intelligent fusion of the information provided by both sensors could reduce false alarms and the advent of non detected pedestrians, thereby increasing the performance of a pedestrian detection and surveillance system.
Approach: The detection of pedestrians is a process involving several interdependent steps. The quality of the steps involving data acquisition, locating zones of movement, classification and monitoring over time is crucial for a more robust detection. Data acquisition requires the constitution of a database which combines sequences of visible and infrared images obtained under difference climatic and lighting conditions. The sequence of images must be synchronized, corrected and calibrated both geometrically and with temperature. The extraction and monitoring over time of each region of interest makes use of movement and is carried out independently for each sequence. A matching of the regions of interest is developed using the epipolar constraints. Finally, for the step involving the classification, critical parameters indicating the presence of people are determined on the basis of characteristics such as temperature, geometry and ratios compared to the rest of the environment.
Challenges: The detection and monitoring of people in interior and exterior environments involves numerous challenges. Algorithms treating the detection of people already exist in the Computer Vision and Systems Laboratory and perform well for visible images (extraction of regions of interest, geometric calibration). One of the challenges is to adapt these algorithms for the treatment of infrared images. Then, the respective limitations of the two sensors must be clearly identified so as to extract the complementary information. The greatest challenge involves the development and proposal of a method of intelligent fusion which will enable the robustness of human detection to be improved while reducing false alarms and the advent of non detected pedestrians.
Applications: The applications of a visible sensor for pedestrian detection and monitoring are already numerous and can be applied to many public environments (parking lots, airports, etc.). With the addition of an infrared sensor, these systems will become more robust and will be able to function under varying lighting and climatic conditions, both day and night, in summer as well as in winter.
Calendar: September 2002 - December 2004
Last modification: 2007/09/28 by torresan


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