Marques, Oge

There is a substantial amount of evidence that suggests that driver drowsiness
plays a significant role in road accidents. Alarming recent statistics are raising the
interest in equipping vehicles with driver drowsiness detection systems. This dissertation describes the design and implementation of a driver drowsiness detection system that is based on the analysis of visual input consisting of the driver's face and eyes. The resulting system combines off-the-shelf software components for face detection, human skin color detection and eye state classification in a novel way. It follows a behavioral methodology by performing a non-invasive monitoring of external cues describing a driver's level of drowsiness. We look at this complex problem from a
systems engineering point of view in order to go from a proof-of-concept prototype to
a stable software framework. Our system utilizes two detection and analysis methods:
(i) face detection with eye region extrapolation and (ii) eye state classification.
Additionally, we use two confirmation processes - one based on custom skin color
detection, the other based on nod detection - to make the system more robust and
resilient while not sacrificing speed significantly. The system was designed to be dynamic and adaptable to conform to the current conditions and hardware capabilities.

In this dissertation we apply sparse constraints to improve optical flow and
trajectories. We apply sparsity in two ways. First, with 2-frame optical flow, we
enforce a sparse representation of flow patches using a learned overcomplete dictionary. Second, we apply a low rank constraint to trajectories via robust coupling. We begin with a review of optical flow fundamentals. We discuss the commonly used flow estimation strategies and the advantages and shortcomings of each. We introduce the concepts associated with sparsity including dictionaries and low rank matrices.

Intrusion Detection Systems (IDS) are security tools which monitor systems and networks for malicious activity. In saturated network links the amount of data present for analysis can overwhelm them, resulting in potentially undetected attacks. Many of these network links contain significant amounts of multimedia traffic which may seem to contribute to the problem, however our work suggests otherwise. This thesis proposes a novel method to classify and analyze multimedia traffic in an effort to maximize the efficiency of IDS. By embedding multimedia-specific knowledge into IDS, trusted multimedia contents can be identified and allowed to bypass the detection engine, thereby allowing IDS to focus its limited resources on other traffic. The proposed framework also enables IDS to detect multimedia-specific exploits which would otherwise pass under the radar. Results of our experiments confirm our claims and show substantial CPU savings in both streaming and non-streaming scenarios.

Academic advising is an important and time-consuming task and different tools and techniques can be used to make it an effective and efficient process. This thesis describes the design and development of a Web-based advising system that supplements the conventional advising process. The goals of the system include: to minimize repetitive tasks performed by advisors, to encourage students to adopt a proactive attitude towards advising, to make advising-related information available to remote students in a single place, in electronic format, and to minimize inconsistencies in the advising process. The system supports three different types of users (students, advisors, and secretaries). This thesis proposes a new Web-based advising system model. It also presents its architecture and an implementation of a prototype. Web-based advising system introduces a new approach towards advising over the Internet. Lessons learned from various experiments of the prototype are discussed in this thesis.

Object segmentation in a video sequence is an essential task in video processing and forms the foundation of content analysis, scene understanding, object-based video encoding (e.g. MPEG-4), various surveillance and 2D-to-pseudo-3D conversion applications. Popularization and availability of video sequences with increased spatial resolution requires development of new, more efficient algorithms for object detection and segmentation. This dissertation discusses a novel neural-network-based approach to background modeling for motion-based object segmentation in video sequences. In particular, we show how Probabilistic Neural Network (PNN) architecture can be extended to form an unsupervised Bayesian classifier for the domain of video object segmentation. The constructed Background Modeling Neural Network (BNN) is capable of efficiently handling segmentation in natural-scene sequences with complex background motion and changes in illumination. The weights of the proposed neural network serve as an exclusive model of the background and are temporally updated to reflect the observed background statistics. The proposed approach is designed to enable an efficient, highly-parallelized hardware implementation. Such a system would be able to achieve real-time segmentation of high-resolution image sequences.

This dissertation presents the results of research that led to the development of a complete, fully functional, image search and retrieval system with relevance feedback capabilities, called MUSE (MUltimedia SEarch and Retrieval Using Relevance Feedback). Two different models for searching for a target image using relevance feedback have been proposed, implemented, and tested. The first model uses a color-based feature vector and employs a Bayesian learning algorithm that updates the probability of each image in the database being the target based on the user's actions. The second model uses cluster analysis techniques, a combination of color-, texture-, and edge(shape)-based features, and a novel approach to learning the user's goals and the relevance of each feature for a particular search. Both models follow a purely content-based image retrieval paradigm. The search process is based exclusively on image contents automatically extracted during the (off-line) feature extraction stage. Moreover, they minimize the number and complexity of required user's actions, in contrast with the complexity of the underlying search and retrieval engine. Results of experiments show that both models exhibit good performance for moderate-size, unconstrained databases and that a combination of the two outperforms any of them individually, which is encouraging. In the process of developing this dissertation, we also implemented and tested several image features and similarity measurement combinations. The result of these tests---performed under the query-by-example (QBE) paradigm---served as a reference in the choice of which features to use in the relevance feedback mode and confirmed the difficulty in encoding the understanding of image similarity into a combination of features and distances without human assistance. Most of the code written during the development of this dissertation has been encapsulated into a multifunctional prototype that combines image searching (with or without an example), browsing, and viewing capabilities and serves as a framework for future research in the subject.