One important application of image processing and computer vision is traffic monitoring and control. In this thesis, a system for detection of moving vehicles approaching an intersection from a sequence of color images acquired by a stationary camera in the context of traffic light control systems is presented. As the system is dedicated to outdoor applications, efficient and robust vehicle detection under various weather and illumination conditions is examined. To deal with these ever changing conditions, vehicle detection relies on motion segmentation and on an original algorithm using color mapping to achieve feature space segmentation. Experimental results on real outdoor sequences of images demonstrate the system's robustness under various environmental conditions.

Electronic commerce is growing at an ever increasing pace every year. However, the payment mechanisms in use were developed without electronic commerce in mind, and as a result, must be modified before they may be used for online transactions. In this thesis, the online payment mechanisms in use are analyzed to find out if they meet present day electronic businesses requirements. A comparative study of existing payment mechanisms has been performed and their strengths and weaknesses evaluated. Recommendations are made for improved mechanisms capable of meeting the needs of consumers and businesses today.

Acceptance sampling is an important field of statistical quality control. This thesis proposed online algorithms to construct hypergeometric sampling plans using Chebyshev polynomials. The recurrence relationship of Chebyshev polynomials significantly reduces the combinatorial computations for hypergeometric distributions. This property can be extended to compute binomial probabilities. The online tables generated by the proposed algorithms are easy to use and precise compared with the Odeh and Owen (1983)'s table. One can readily check the required sample size and acceptance number for pre-specified producer's risk; consumer's risk, proportion defective, and/or lot size for hypergeometric and binomial samplings. The algorithms also cover the confidence limits for hypergeometric and OC curve for binomial sampling. The programs written in R language are appended to show the implementation for the algorithms proposed.

Metadata is descriptive information for data. The purpose of metadata is to facilitate describing, managing and discovering resources in huge distributed repositories. Metadata experts worldwide create the Dublin Core (DC), which acts as a fundamental core metadata standard on which industrial metadata standards based. For educational industry, the need is increasing for description and exploration of a learning object (LO) in distributed learning object repositories (LOR) worldwide. Several organizations aim to establish metadata standards for facilitating better identifying, exchanging and reusing learning objects according to their specific needs. The Institute of Electrical and Electronics Engineers (IEEE) published Learning Object Metadata (LOM), which is a credited standard on the global level since it best represents the characteristics of digital learning objects. Conversion from one metadata standard to another is necessary in case that people want to exchange and reuse learning objects tagged using different kinds of metadata standards. Mapping between the DC and the LOM is an essential job in many e-learning systems. In this thesis, we present a new web based metadata editor for the DC and LOM, and a Web Services oriented mapping tool between them. Other clients can use our editor to create DC or LOM metadata records when catalog their learning objects into our LOR, and integrate the mapping web services as a part of different systems regardless of different platforms, protocols, and displaying devices. Our objective is to promote the reusability and interoperability for both the DC and the LOM users, therefore benefit the learning object industry by lowering the cost of using metadata. The DC-LOM mapping tool is demonstrated in our e-learning system called UbiLearn developed at the MCRLab of University of Ottawa.

In a multicast network of a large number of receivers where transmission resources and end-systems are of high heterogeneity, the receivers require automatically selected heterogeneous QoS stream variants that meet their end-system requirements. In this thesis, we introduce a Receiver Satisfaction Driven (RSD) method that takes into account user preferences stored in user profiles in order to automatically select the stream variants that maximize the overall user satisfaction. In the case of a large number of receivers, the receivers may be partitioned into different user classes according to their bandwidth limit. Each class is allowed to send one or several representative group profiles, depending on the number of users in that group. The source processes the data from the group profiles and determines the optimal QoS parameters for a given number of stream variants. Our simulation results show that the streams variants generated based on representative group profiles give a reasonable satisfaction level for all receivers.

Artificial neural networks are frequently used in business, engineering, and scientific applications. Their increasing popularity is due to the satisfactory results achieved by using artificial neural networks applications for forecasting. The aim of the thesis is to carry out a comparative analysis of the forecasting performance of different neural network models. The initial portion of the thesis will entail a study of the background as well as the different methodologies of artificial neural networks. Focus will be given to time series forecasting models. Results of forecasting different sets of data show that using neural network models gives almost the same, and in some cases better results, than other forecasting methods (Paplinski, 2003). Throughout this thesis paper, I intend to provide an introduction to neural networks, and the models applicable to time series forecasting. Particularly, in the feed-forward model of neural networks, I will provide some examples using R statistical programming language that will help the reader to observe a practical usage of R as a software that helps the user to compute complicated neural networking problems.

This thesis studies the cycle-based protection schemes for DWDM (Dense Wavelength Division Multiplexing) networks. We first present the MSCC (Min-Sum Cycle Cover) Algorithm for simultaneous detection and protection using the Eulerian graph theory, and then analyze the performance of the algorithm through making comparison of it with other algorithms. By extending the features of algorithm SLA (Straddling Link Algorithm) and initial ideas for p-cycle candidate generation (operations called Add and Join), we formulate more advanced cycle-generation algorithms: SP-Add, SP-Join and SP-Merge. We then use the developed heuristic spare capacity planning algorithms based on weighted or unity capacity efficiency of p-cycles, WCIDA, to test the fully restorable p-cycle network design. In addition, two capacitated algorithms will be presented and compared with each other.

This thesis is targeted at proposing a framework for building a voice-enabled web application which allows users to interact with the system through the Internet, using either speech (microphone, speaker) or traditional devices (such as keyboard or mouse). Several voice-related technologies are briefly described in the beginning of this thesis. The original design of the Click&Dial system was modified and enhanced in order to adopt a popular J2EE framework named Struts. Compared with SALT, XHTML plus Voice and Voice XML, SALT is chosen as the main technology for building the speech-enabled web application. The architecture of integrating voice into the web application is presented, and the framework is implemented. With the help of it, the speech-based components are demonstrated to be embedded easily into the system, and the simplicity of the migration is guaranteed as well. The proposed system, with both visual interface and voice interface, is designed and developed using Java-based technology combined with embedded speech engines. The implementation successfully verifies that the proposed framework meets the design requirements and is capable of supporting two kinds of user interfaces for a web application, and the targeted objective is achieved. Additionally, a possible solution for voice-activated Click&Dial system on Pocket PCs is recommended, which enables the voice-enabled application to be accessed easily from anywhere using hand-held devices in the future.

This work investigates transfer pricing models among network-manufacturing firms for a distributed product. The particularities of the models are: instead of each company trying to maximize individually the value added to its supply chain in which it is embedded, The models propose to maximize the value added by the network-companies in the global supply chain. Under specific assumptions on the nature of production, cost and value functions in typical production/distribution companies, it optimizes the supply chain structure for network-companies, distributed in one economic region. We calculate the transfer price and share the value added between the networked firms by two approaches: resource-based approach and efficiency-based approach. The models are formally defined, optimally formulated, and solved.

In this thesis, we systematically examine the ECN (Explicit Congestion Notification) mechanism to be applied to some AQM (Active Queue Management) techniques developed in recent years, namely, RED (Random Early Detection), BLUE, ARED (Adaptive RED) and PI-RED (Proportional and Integral RED). We use packet marking as the indication of network congestion, instead of the strategy of dropping packet currently adopted by these AQM algorithms. The performance comparison of ECN-AQM with AQM has been conducted in not only a single bottleneck network but also a multi-bottleneck network environment. The simulation results have shown that ECN-AQM algorithms can greatly decrease unnecessary packet loss, one of the main shortcomings of RED and some other AQM algorithms. They can also reduce average queue size in most cases, but the queue size oscillation problem remains. We therefore investigate the application of the head dropping policy to AQM techniques, as a simple solution to the problem of queue oscillation, another shortcoming of RED and its variants. With this method, instead of tail dropping, which is currently used by RED and many other AQM schemes, the TCP source can be informed of the congestion occurring in the bottleneck router earlier by getting rid of time to wait through the queuing delay. We have compared DH-RED (Drop Head RED) and DH-BLUE (Drop Head BLUE) with the current RED and BLUE in both of the single bottleneck and the multi-bottleneck networks. We found the performance of queue size stability can be greatly improved by DH-RED and DH-BLUE.