Design and implementation of real time computer vision algorithms for video surveillance applications

Abstract

In this dissertation, we present our research work held at the Center of Robotics (CAOR) of the Ecole des Mines de Paris which tackles the problem of intelligent video analysis. The primary objective of our research is to prototype a generic framework for intelligent video analysis. We optimized this framework and configured it to cope with specific application requirements. We consider a people tracker application extracted from the PUVAME project. This application aims to improve people security in urban zones near to bus stations. Then, we have improved the generic framework for video analysis mainly for background subtraction and visual object detection. We have developed a library for machine learning specialized in boosting for visual object detection called LibAdaBoost. To the best of our knowledge LibAdaBoost is the first library in its kind. We make LibAdaBoost available for the machine learning community under the LGPL license. Finally we wanted to adapt the visual object detection algorithm based on boosting so that it could run on the graphics hardware. To the best of our knowledge we were the first to implement visual object detection with sliding technique on the graphics hardware. The results were promising and the prototype performed three to nine times better than the CPU. The framework was successfully implemented and integrated to the RTMaps environment. It was evaluated at the final session of the project PUVAME and demonstrated its fiability over various test scenarios elaborated specifically for the PUVAME project.