Computer vision algorithms have played a vital role in video surveillance systems to detect surveillance events for public safety and security. Even so, a common demerit among these systems is their unfitness to handle divers crowded scenes. In this thesis, we have developed algorithms which accommodate some of the challenges encountered in videos of crowded environments (e.g., airports, malls, sporting events) to a certain degree. We have adopted approaches by first performing a global-level motion analysis within each frame's region of interest that provides the knowledge of crowd's multi-modal behaviors in the form of complex spatiotemporal structures. These structures are then employed in the detection of unusual surveillance events occurred in the crowds. To conduct experiments, we have heavily relied on three thought-provoking datasets. The results reflect some unique global excellences of the approaches. We have also developed a pseudo Euclidian distance. To show its usage, a methodology based on it has been employed in the detection of various usual surveillance events from the TRECVID2008. Some results report the robustness of the methodology, while the rest gives evidence of the difficulty of the problem at hand. Big challenges include, but are not limited to, massive population flow, heavy occlusion, reflection, shadow, fluctuation, varying target sizes, etc. Notwithstanding, we have got much useful insights and experience to the practical problems. In addition, the thesis explores an individual target tracking algorithm within miscellaneous crowded scenes. Video sequences from the PETS2009 Benchmark data have been used to evaluate its performance. Viewing its pros and cons, the algorithm is still highly accurate and its sensitivity to the effects of diversity in noise and lighting, which ascertains its high-quality performance on disappearances, targets moving in and out of the shadow, and flashes of light.
Directeur de thèse : Chabane DJERABA, (Professeur) Université de Lille I Rapporteurs : Zhongfei (Mark) ZHANG, (Professeur) State University of New York (SUNY) Claude CHRISMENT, (Professeur) Université de Toulouse III Membres : Sophie TISON, (Professeur) Université de Lille I Liming CHEN, (Professeur) Ecole Centrale de Lyon Bernard GOSSELIN, (Professeur) Université de Mons
Thesis of the team FOX defended on 16/07/2010