Adaptive algorithms for background estimation to detect moving objects in videos. (Algorithmes adaptatifs d'estimation du fond pour la détection des objets mobiles dans les séquences vidéos)

Detecting foreground pixels is the rst step to detect objects of interestin videos. The objective of this thesis is to propose a new background estimationmethod to detect foreground pixels. The proposed method can adapt the estimatedbackground to various changes of environment (e.g. changes of illumination or ofcontextual objects).The proposed background estimation method consists of a new background sub-traction algorithm to detect foreground pixels, post-processing algorithms to removeshadow and highlight, and a controller to adapt the background subtraction algo-rithm to the current scene conditions.The new background subtraction algorithm takes into account the scene char-acteristics such as dynamic background (e.g. tree leave motion), displacement ofcontextual objects to improve the foreground detection results. It also proposes anew updating method to better adapt its background representation to the currentscene conditions.The algorithms to remove shadow and highlight employ new chromaticity andhomogeneity (texture) constraints which are robust to illumination changes. Theseconstraints are constructed based on the illumination model and the camera model.The controller has two adaptation methods for the background subtraction al-gorithm. The rst method is to selectively update the background representation ofthe background subtraction algorithm. With this updating method, the backgroundsubtraction algorithm can solve various problems such as managing stationary ob-jects, keeping track of objects when they stop moving. The second method is totune the parameter values of the background subtraction algorithm. To ful ll thesetwo tasks, the controller extensively uses the feedback from the classi cation taskand the information about the background subtraction algorithm and the scene.This method has been validated using the public database ETISEO and onehour video from the project GERHOME.