Video object segmentation and tracking.

Abstract

One of the more complex video processing problems currently vexing researchers is that of object segmentation. This involves identifying semantically meaningful objects in a scene and separating them from the background. While the human visual system is capable of performing this task with minimal effort, development and research in machine vision is yet to yield techniques that perform the task as effectively and efficiently. The problem is not only difficult due to the complexity of the mechanisms involved but also because it is an ill-posed problem. No unique segmentation of a scene exists as what is of interest as a segmented object depends very much on the application and the scene content. In most situations a priori knowledge of the nature of the problem is required, often depending on the specific application in which the segmentation tool is to be used. This research presents an automatic method of segmenting objects from a video sequence. The intent is to extract and maintain both the shape and contour information as the object changes dynamically over time in the sequence. A priori information is incorporated by requesting the user to tune a set of input parameters prior to execution of the algorithm. Motion is used as a semantic for video object extraction subject to the assumption that there is only one moving object in the scene and the only motion in the video sequence is that of the object of interest. It is further assumed that there is constant illumination and no occlusion of the object. A change detection mask is used to detect the moving object followed by morphological operators to refine the result. The change detection mask yields a model of the moving components; this is then compared to a contour map of the frame to extract a more accurate contour of the moving object and this is then used to extract the object of interest itself. Since the video object is moving as the sequence progresses, it is necessary to update the object over time. To accomplish this, an object tracker has been implemented based on the Hausdorff objectmatching algorithm. The dissertation begins with an overview of segmentation techniques and a discussion of the approach used in this research. This is followed by a detailed description of the algorithm covering initial segmentation, object tracking across frames and video object extraction. Finally, the semantic object extraction results for a variety of video sequences are presented and evaluated.