Dominant object detection for autonomous vision-based surveillance

The deployment of visual surveillance and monitoring systems has reached massive proportions. Consequently, a need to automate the processes involved in retrieving useful information from surveillance videos, such as detecting and counting objects, and interpreting their individual and joint behavior, has emerged. The existing methods targeting such automation and working towards "smart" surveillance solutions largely rely on pattern recognition solutions trained offline in a supervised fashion. In this thesis we address the challenge of minimizing the level of supervision in automating the smart surveillance mechanisms. In particular, we focus on the development of fully autonomous detectors of objects that are typical for the observed scene. Using such solutions, objects can be observed and counted (e.g. pedestrians in a shopping street, cars and trucks on a highway), but also atypical objects can be detected leading to automatic security alerts. The rationale behind our proposed solution is that the characteristics of a scene or objects involved can be learned if sufficient time for observation is given, and if the measured features characterizing the properties of the scene and objects are generic enough. In the first step, we demonstrate how candidates for the typical objects can be separated from the scene background using an existing robust motion segmentation mechanism and our novel shadow removal method. In the second step, the set of candidate objects is filtered using an automatically learned perspective deformation model to remove irrelevant objects and compensate for the imperfections of the motion segmentation and shadow removal processes. Finally, in the third step, a module is developed, in which the candidate typical objects are used as positive training samples to train an advanced typical object detector. The effectiveness of this three-step approach is demonstrated first on the case of a single typical object class present in the observed scene. Then, we also show how the proposed approach can be expanded to simultaneously detect two such object classes, if present