Abstract

Segmentation of objects from a video sequence is one of the basic operations commonly employed in vision-based systems. The quality of the segmented object has a profound effect on the performance of such systems. Segmentation of an object becomes a challenging problem in situations in which the background scenes of a video sequence are not static or contain the cast shadow of the object. This thesis is concerned with developing cost-effective methods for object segmentation from video sequences having dynamic background and cast shadows.
A novel technique for the segmentation of foreground from video sequences with a dynamic background is developed. The segmentation problem is treated as a problem of classifying the foreground and background pixels of the frames of a sequence using the pixel color components as multiple features of the images. The individual features representing the pixel gray levels, hue and saturation levels are first extracted and then linearly recombined with suitable weights to form a scalar-valued feature image. Multiple features incorporated into this scalar-valued feature image allows to devise a simple classification scheme in the framework of a support vector machine classifier. Unlike some other data classification approaches for foreground segmentation, in which a priori knowledge of the shape and size of the moving foreground is essential, in the proposed method, training samples are obtained in an automated manner. The proposed technique is shown not to be limited by the number, patterns or dimensions of the objects.
The foreground of a video frame is the region of the frame that contains the object as well as its cast shadow. A process of object segmentation generally results in segmenting the entire foreground. Thus, shadow removal from the segmented foreground is essential for object segmentation. A novel computationally efficient shadow removal technique based on multiple features is proposed. Multiple object masks, each based on a single feature, are constructed and merged together to form a single object mask. The main idea of the proposed technique is that an object pixel is less likely to be indistinguishable from the shadow pixels simultaneously with respect to all the features used.
Extensive simulations are performed by applying the proposed and some existing techniques to challenging video sequences for object segmentation and shadow removal. The subjective and objective results demonstrate the effectiveness and superiority of the schemes developed in this thesis.