Developing object detection, tracking and image mosaicing algorithms for visual surveillance

Official URL: http://risc01.sabanciuniv.edu/record=b1534381 (Table of Contents)

Visual surveillance systems are becoming increasingly important in the last decades due to proliferation of cameras. These systems have been widely used in scientific, commercial and end-user applications where they can store, extract and infer huge amount of information automatically without human help. In this thesis, we focus on developing object detection, tracking and image mosaicing algorithms for a visual surveillance system. First, we review some real-time object detection algorithms that exploit motion cue and enhance one of them that is suitable for use in dynamic scenes. This algorithm adopts a nonparametric probabilistic model over the whole image and exploits pixel adjacencies to detect foreground regions under even small baseline motion. Then we develop a multiple object tracking algorithm which utilizes this algorithm as its detection step. The algorithm analyzes multiple object interactions in a probabilistic framework using virtual shells to track objects in case of severe occlusions. The final part of the thesis is devoted to an image mosaicing algorithm that stitches ordered images to create a large and visually attractive mosaic for large sequence of images. The proposed mosaicing method eliminates nonlinear optimization techniques with the capability of real-time operation on large datasets. Experimental results show that developed algorithms work quite successfully in dynamic and cluttered environments with real-time performance.