Degree Name

Doctor of Philosophy


School of Electrical, Computer and Telecommunications Engineering, Faculty of Informatics


Visual object tracking has been extensively investigated in the last two decades for its attractiveness and profitability. It remains an active area of research because of the lack of a satisfactory holistic tracking system that can deal with intrinsic and extrinsic distortions. Illumination variations, occlusions, noise and errors in object matching and classification are only afraction oftheproblems currently encountered in visual object tracking. The work developed in this thesis integrates contextual information in a Bayesian framework for object tracking and abnormal behavior detection; more precisely, it focuses on the intrinsic characteristics of video signals in conjunction with object behavior to improve tracking outcomes. Therepresentationofprobabilitydensityfunctionsisessentialformodelingstochas¬tic variables. In particular, parametric modeling is convenient since it makes possi¬ble the efficient storage of the representation and the simulation of the underlying stochastic process. The Gaussian mixture model is employed in this thesis to rep¬resent the pixel color distribution for segregation of foreground from background. The model adapts quickly to fast changes in illumination and resolves the problem of “pixel saturation” experienced by some existing background subtraction algo¬rithms. The technique leads to better accuracy in the extraction of the foreground for higher-level tasks such as motion estimation. The solutionof theBayesianinferenceproblemforMarkovchainsand,inparticular, the well-known Kalman and particle filters is also investigated. The integration of contextual inference is of paramount importance in the aforementioned estimators; it resultsinobject-specifictracking solutionswithimproved robustness. Thevehicle tracking problem is explored in detail. The projective transformation, imposed by the environment configuration, is integrated into the Kalman and particle filters, which yields the “projective Kalman filter” and the “projective particle filer”. Ex¬tensive experimental results are presented, which demonstrate that the projective Kalman and particle filters improve tracking robustness by reducing tracking drift and errors in the estimated trajectory. The constraint on the known nature of the environment is then relaxed to allow general tracking of pedestrians. A mixture of GaussianMarkovrandom fieldsisintroduced tolearnpatternsof motionand model contextualinformationwithparticle filtering. Suchinferenceresultsinanincreased tracking robustness to occlusions. The local modeling with the Markov random fields also provides inference on ab-normalbehaviordetection. Sincelocalpatterns are unveiledby theMarkov random field mixture, detecting abnormal behavior is reduced to the matching of an object feature vector to the underlying local distribution, whereas the global approach, introducinggeneralization errors, involves complex, cumbersome andinaccuratede¬cisions. Experimental evaluationonsyntheticand realdatashow superiorresultsin abnormal behavior detection for driving under the influence of alcohol and pedes¬trians crossing highways.

02Whole.pdf (6115 kB)