Computer Science

Probabilistic Trans-Algorithmic Search

Bilal Gonen

Dr. Murat Yuksel

Monday, June 8th, 2009 at 09:30a.m.

Computer Networking Lab (Next to AGN, SEM201)

Real large-scale systems exhibit varying behavior and give time-dependent responses due to internal failures and changes. For online optimization of such large-scale systems, a single black-box optimization algorithm is not capable of handling their varying behavior. Some of the search algorithms may perform better than other search algorithms on the same black-box problem depending on the response surface. In response to an internal failure, the administrators have to reconfigure the system in a very short period of time. Most of the real systems require parameter (re)configuration to be done within a limited amount of time, and thus arrival to a complete but not necessarily optimal solution is of vital important for large-scale systems with inevitable internal failures or changes. This parameter (re)configuration problem is even more complicated for networked systems due to their distributed nature. We propose a Trans-Algorithmic Search (TAS) framework which leverages multiple optimization search algorithms in an iterative manner to find a better parameter vector for the black-box system-at-hand. The TAS framework aims to automatically find out the best set of algorithms that should work on the problem-at-hand. In this sense, TAS applies an evolutionary algorithms to “search for the best search”. TAS allocates experiment budget to each available search algorithm and observes the success of them on the problem. Depending on their successes, TAS probabilistically reallocates the experiment budget for the next round by using a roulette wheel approach. Following each round, the TAS framework allows "transfer" of best found results among the algorithms being used, which makes our framework a "trans-algorithmic" one. We compare TAS against well-known black-box algorithms such as Simulated Annealing and Genetic Algorithm, and show that it can outperform most of the search algorithms on well-known benchmark objective functions.

Computer Science - Software Engineering

Towards Automatic Parallel Game Architectures

Dave Carr

Eelke Folmer

Tuesday, May 5th, 2009 at 1:00p.m.

AGN (SEM-201)

As video games steadily increase in complexity and detail, game engines must also improve to be able to support both rapid and modular development while still maintaining high performance. Component-based architectures have been shown to be effective for allowing developers to rapidly create stunning games with modular components, but unfortunately come with a performance cost over traditional call-and-return systems. This thesis proposes a method that modifies existing component-based game engine architectures to automatically distribute and synchronize work in order to improve performance.

Computer Science and Engineering

BIOMETRIC WATERMARKING OF 3D MULTIMEDIA

Rakhi Motwani

Dr. Fred Harris @ 784-6571 (Fred.Harris@cse.unr.edu)

Tuesday, May 5th, 2009 at 09:30a.m.

Access Grid Node, Scrugham Engineering and Mines (SEM 201)

For the past few years, 3D graphics has been gaining a lot of popularity. Online 3D warehouses, evolution of online social networks from 2D spaces (chats, imaging, video sharing) to 3D domain (avatars in 3D virtual worlds like Second Life), and 3D online game environments have created a strong need to provide artists with new weapons in their battle against online pirates. Copyright issues are crucial in media arts to enforce ownership rights in the on-line world since digital artwork can easily be duplicated, altered, and plagiarized. Watermarking techniques have been used for copyright protection of digital media. Watermarking is about using steganography to encode small information into multimedia with no perceivable difference. Traditional watermark- ing techniques embed text, cryptographic keys, copyright ownership messages, logos, and image or digital media content-based information into the multimedia. This project takes existing wa- termarking systems to the next level by inserting biometric data into the multimedia. Biometrics is used to enhance security levels of the watermarking techniques. This application implements a watermarking scheme that uses the artists voice to create a unique watermark for insertion into the 3D graphics model. The watermark is generated by using a Gaussian Mixture Model (GMM) representation of an artists speech. Melfrequency cepstral coefficients, representing feature vectors of the sound signal, are used to create the GMM. Benefit of using voice, as a biometric, is that it is easy to capture, since it requires no special equipment (as opposed to retina or iris scanners, fingerprint capturing devices).

Computer Science and Engineering

Minimizing Multi-Hop Wireless Routing State under Application-Based Accuracy Constraints

Mustafa Omer Kilavuz

Dr. Murat Yuksel

Monday, May 4th, 2009 at 11:00a.m.

SEM 331

Provisioning of rich routing building blocks to mobile ad-hoc networking applications has been of high interest. Several MANET applications need flexibility in describing paths their traffic will follow. To accommodate this need, previous work has proposed several viable routing schemes such as Dynamic Source Routing (DSR) and Trajectory-Based Routing (TBR). However, tradeoffs involved in the interaction of these routing schemes and the application-specific requirements have not been explored. Especially, techniques to help the application to do the right routing choices are much needed. We consider techniques that minimize routing protocol state costs under application-based constraints. We study the constraint of “accuracy” of the application’s desired route, as this constraint provides a range of choices to the applications. As a crucial part of this optimization framework, we investigate the tradeoff between the packet header size and the network state. We, then, apply our frame work to the case of TBR with application-based accuracy constraints in obeying a given trajectory. We begin with simple discrete models to clarify the tradeoff between packet header size and network state. We show that the problem of accurate representation of a trajectory with the objective of minimizing the cost incurred due to header size and network state is difficult to solve optimally. Finally, we develop heuristics solving this problem and illustrate their performance.

Computer Science and Engineering.

THIRD GENERATION 3D MODEL WATERMARKING

Mukesh Motwani

Dr. Fred Harris @ 784-6571 (Fred.Harris@cse.unr.edu)

Monday, May 4th, 2009 at 10:00a.m.

Access Grid Node, Scrugham Engineering and Mines (SEM 201)

With the explosion of multimedia content over Internet, there is a need for copyright protection of digital content. Whether it is music albums swapped over peer to peer networks or video files uploaded over YouTube or 3D models such as Shrek, artists need to protect their ownership of content. Multimedia Watermarking provides a solution to piracy of multimedia content by embedding a hidden piece of information in the original content. Typically, a logo as an image, or a random string or a number is inserted as watermark. It is required that the watermark should survive unintentional attacks such as compression loss, affine transformations and noise. At the same time, the watermark should be perceptually invisible and should be embedded in such a way so that the potential hacker should be forced to make substantial changes in the multimedia content in order to destroy the watermark. Watermarking algorithms have a basic requirement that the watermark amplitude should be as high as possible for robustness and at the same time the watermark should not introduce any perceptible artifacts. Thus, the design of watermarking algorithms involves a trade off between imperceptibility and robustness. The first generation of 3D watermarking techniques would insert the watermark based on spatial geometry. The second generation of 3D watermarking explored the use of multi-resolution transform such as wavelet representation to insert the watermark and improve the robustness. The amount of information inserted as watermark acts as the noise to the host signal or the 3D model. Shannons Information theory puts theoretical limit on the Signal to Noise ratio for watermark insertion. This dissertation explores the use of AI techniques to insert high density watermarks and go the extra mile in terms of hiding more information than the first and second generation techniques. The focus of this study is to maximize the energy of the watermark and “intelligently” selecting vertices in 3D model for watermark insertion and at the same time still maintaining randomness in the process. Use of Artificial Neural Networks and Fuzzy Logic has been proposed in the study.

Computer Science and Engineering

Towards Generalized Accessibility of Video Games for the Visually Impaired

Bei Yuan

Dr. Fred Harris @ 784-6571 (Fred.Harris@cse.unr.edu) Dr. Eelke Folmer @ 784-7952 (efolmer@cse.unr.edu)

Wednesday, April 29th, 2009 at 11:00a.m.

Scrugham Engineering and Mines (SEM) 201

Over the last three decades, video games have evolved from an obscure pastime to a force of change that is transforming the way people perceive, learn about, and interact with the world around them. Video games are not only a popular form of entertainment, but are increasingly being used for other purposes, such as education and health, as well. Despite this increased interest, a significant number of people encounter barriers when playing games, due to a disability.
This dissertation helps provide an understanding of how video games can be designed and modi- fied to improve their accessibility features. A large number of existing, accessible games have been studied and analyzed to provide us with insights and understanding as to the importance of encouraging universal access in this field. Though our survey work covered several types of disabilities, the bulk of this dissertation focuses on improving accessibility for the visually impaired. Specific design strategies are illustrated and proven by the development and evaluation of actual blind-accessible games.
Case studies are presented for each of the three games we developed during the research period. The first two case studies introduce new strategies for developing blind-accessible games and improv- ing their current state-of-the-art. The third demonstrates a new game for sighted users, the result of playing which, is improved accessibility in another existing game (Second Life). Furthermore, user studies were conducted that focused on the enjoyment, educational, and social interaction aspects of these games while evaluating their ease of access.

Computer Science and Engineering

Mind Reading for Social Robots: Stochastic Models of Intent Recognition

Richard Kelley

Dr. Monica Nicolescu

Thursday, April 23rd, 2009 at 12:30a.m.

SEM 201

Understanding intent is an important aspect of communication among people and is an essential component of the human cognitive system. This capability is particularly relevant for situations that involve collaboration among multiple agents or detection of situations that can pose a particular threat. To build robots that reliably function in the human social world, we must develop models that robots can use to mimic the intent recognition skills found in humans. In this work, we propose an approach that allows a physical robot to detect the intentions of others based on experience acquired through its own sensory-motor capabilities, later using this experience while taking the perspective of the agent whose intent should be recognized. Our method uses Hidden Markov Models (HMMs) designed to model a robot`s experience and interaction with the world when performing various actions. We augment this baseline intent recognition system with a framework that supports the use of contextual information to improve the overall system`s performance. Additionally, we explore the use of an evolutionary algorithm for solving the decoding problem for generative stochastic models, in our intent recognition application and in other domains. We validate all of our approaches on physical robots that classify the intentions of several people performing various activities in a number of scenarios.

Computer Science & Engineering

A Non-Parametric Framework for Object Tracking in Videos with Quasi-Stationary Backgrounds

Alireza Tavakkoli

Dr. Mircea Nicolescu (775) 784-4356 mircea@cse.unr.edu

Wednesday, April 22nd, 2009 at 02:30a.m.

SEM 201

The ability to automatically detect and track objects of interest in a video sequence is an essential feature of many high-level vision-based applications. In this dissertation we propose a non-parametric computational framework which detects and tracks foreground image regions (typically corresponding to moving objects such as people or vehicles), while estimating their trajectories for further processing stages such as activity recognition. The contribution of this work extends along two main directions: background modeling for object detection and appearance-based object tracking. In many applications such as vision-based surveillance or traffic monitoring it is typically assumed that the camera is static. Even with this assumption, quasi-stationary backgrounds (that change due to moving tree branches, waving flags, rain or water surfaces) pose significant challenges in detecting and tracking the actual objects of interest while ignoring such changes. Due to the diverse nature of vision-based applications, it has been a main concern for researchers to design a scene-independent system that consistently handles these difficult situations. In the object detection stage, we first propose a novel adaptive statistical method as a base-line system that addresses the issue of scene-independent background modeling. After investigating its performance, we introduce a universal statistical technique which aims to overcome the weaknesses of its predecessor in modeling slow changes in the background. Furthermore, a new analytical technique is proposed that addresses the limitations of statistical techniques which are bound to the probability density estimation accuracy. The performance of each proposed method is studied, while investigating scenarios where each technique leads to better performance. In the object tracking stage, photometric and geometric appearance models are built for the detected objects, then used to predict their locations in subsequent images by employing a novel spatio-spectral tracking technique. The proposed tracker is shown to be particularly robust to local illumination changes, while also being able to detect and resolve the temporary overlapping of tracked objects. We support our claims with extensive experimental results and comparisons between the proposed techniques and other methods for detection and tracking. Finally, we describe a robotic application which successfully employs the proposed vision-based tracking framework in order to infer the intentions of agents in the monitored environment, before their actions are finalized.

Computer Science and Engineering

CAVEMANDER: A Software Platform for Building Command and Control Applications in CAVE

Sermsak Buntha

Dr. Sergiu Dascalu

Thursday, February 26th, 2009 at 1:00p.m.

SEM-201

Command and control systems play a vital role in displaying information about various operational situations, thus helping decision makers to thoroughly understand them and make related decisions in a timely and correct manner. In military operations, pictures of such situations were traditionally displayed on large tactical boards and/or vertical maps. Although for several decades the computer has been used to replace these traditional displays, the pictures of the situations have largely been presented on 2D media only, such as PC monitors and wall screens. Due to several recognized advantages of 3D visualizations, combined with the power of immersion in virtual worlds, we believe that in command and control applications 3D immersive environments such as the CAVE Automatic Virtual Environment (in short, CAVE) could significantly improve the understanding of the situation and the decision-making performance of the commanders. The present PhD topic proposal aims to investigate existing solutions for developing software for CAVE and propose a new approach and related toolset for creating command and control applications in CAVE. CAVEMANDER, the proposed approach and its related suite of supporting software tools, will allow improved development of command and control applications for CAVE and will consist of a generic software engineering methodology encompassing the necessary construction steps and related artifacts as well as a layered software platform composed primarily of reusable code. CAVEMANDER’s benefits will be assessed on a command and control application in the area of training for naval surface warfare, instantiated in one or more scenarios that will test the trainees’ skills for command and control, including planning and testing, information interpretation, and failure investigation. Although command and control systems are mainly associated with military applications, our research will expand the scope of CAVEMANDER to support applications in other domains, such as fighting wildfires, conducting search and rescue missions, and coordinating planet expeditions. In short, the proposed work is aimed at filling the existing research and development gap in building CAVE-based command and control applications. Future work could encompass improved human-computer interaction solutions, increased reusability, automated code generation, and AI-driven application scenarios.

Computer Science

Wildfire Simulation on the GPU

Roger Viet Hoang

Dr. Fred Harris

Wednesday, November 26th, 2008 at 09:00a.m.

SEM201

The environmental, social, and economic effects of wildfires have led researchers to develop various models to study the behavior of this phenomenon. These models vary widely in terms of complexity, with some models simulating the basic spread of surface fires and other more complex models simulating the progression of fire through tree crowns and lofting embers. The computational requirements of these more complex models limits their use in prediction and interactive applications, and increasing the parallel computational power through the addition of more CPUs is not always cost-effective. At the same time, the increase in computational power of relatively inexpensive graphics cards has led to their use as parallel general purpose processors. This thesis examines the viability of harnessing the power of GPUs to simulate fire spread. A fire spread model that incorporates the effects of surface fire, crown fire, and fire acceleration is developed. A mapping of this model to GPU concepts is presented, and the results of an implementation are discussed.

Computer Science and Engineering

Hand-based Identification/Verification System

Gholamreza Amayeh

Dr. George Bebis

Monday, November 24th, 2008 at 2:45p.m.

SEM201

Hand-based identification/verification is a key biometric technology with a wide range of potential applications both in industry and government. The focus of this work is on improving the efficiency, accuracy, and robustness of hand-based verification. In particular, we propose using high-order Zernike moments to represent hand geometry, avoiding the more difficult and prone to errors process of hand-landmark extraction (e.g., finding finger joints). The proposed system operates on 2D hand silhouette images acquired by placing the hand on a planar lighting table without any guidance pegs, increasing the ease of use compared to conventional systems. Zernike moments are powerful translation, rotation, and scale invariant shape descriptors. To deal with several practical issues related to the computation of high order Zernike moments including computational cost and lack of accuracy due to numerical errors, we have employed an efficient algorithm that uses arbitrary precision arithmetic, a look-up table, and avoids re-computing the same terms multiple times. The focus of this work is on improving the ease of use and accuracy of hand-based authentication systems. Specifically, we employ high-order Zernike moments to represent the segmented parts of the hand silhouette including the palm and the fingers. Segmentation allows us to compensate for finger motion more effectively compared to the initial version of this system that uses Zernike moments of the whole hand. Due to the rotation and translation invariance of the descriptors, the similarity between the query and the templates is computed efficiently using Euclidean distance. The proposed hand-based identification/verification system has been tested on a database of 1000 images from 100 subjects illustrating promising performance. Qualitative comparisons with state of the art systems illustrate that the proposed system has comparable or better performance.

Computer Science and Engineering

An Iterative Multi-Scale Tensor Voting Scheme for Perceptual Grouping of Natural Shapes in Cluttered Backgrounds

Leandro A. Loss

Dr. George Bebis

Friday, November 21st, 2008 at 2:30p.m.

SEM201

Grouping processes organize the data by eliminating irrelevant items and sorting the rest into groups, each corresponding to a particular object. Such organization can provide reliable pre-processed information to higher level computer vision functions, such as object detection and recognition. In our research, we consider the problem of grouping oriented segments in highly cluttered images. To this end, we developed a general and powerful method based on an iterative, multi-scale tensor voting approach. Segments are represented as second-order tensors and communicate with each other through a voting scheme that incorporates the Gestalt principles of visual perception. The key idea of our approach is removing background segments conservatively on an iterative fashion, using multi-scale analysis, and re-voting on the retained segments. We have performed extensive experiments to evaluate the strengths and weaknesses of our approach using both synthetic and real images from publicly available datasets including the fruit-texture dataset and the Berkeley segmentation dataset. Our results and comparisons indicate that the proposed method improves segmentation results considerably, especially under severe background clutter. In particular, we show that using the iterative multi-scale tensor voting approach to post-process the posterior probability map produced by segmentation methods improves boundary detection results in 84% of the grayscale test images in the Berkeley segmentation benchmark. We demonstrate the application of this segmentation approach in high level tasks, such as image vectorization and breast cell detection.

Computer Science and Engineering

A Non-Parametric Framework for Object Tracking in Videos with Quasi-Stationary Backgrounds

Alireza Tavakkoli

Dr. Mircea Nicolescu

Friday, November 21st, 2008 at 11:00a.m.

SEM201

Object tracking is one of the most important tasks in high-level video processing applications. In some cases, such as video surveillance, traffic monitoring, etc., it can be assumed that the camera is static. In such situations the background of the video appears to be stationary. However, some changes in the background such as waving flags, fluctuating monitors, water surfaces, etc. pose significant challenges in detecting and tracking objects of interest. Due to the diverse nature of video applications it has been a difficult to design a scene independent system which models background changes, detects the objects of interest, and robustly tracks them. In this project we use an adaptive statistical method as a base-line system that addresses the issue of scene independent background modeling. We introduce a universal statistical technique which aims to overcome the weaknesses of its predecessor in modeling slow changes in the background. Finally, a new analytical technique is proposed in that approaches the problem of background modeling in a different direction. This technique is introduced in order to solve the limitations of statistical techniques which are bound to the accuracy of the probability density estimation. Once the background of the video is modeled we use the detected objects to feed the object tracking module of the system. This module builds object models for tracking individual objects. The proposed models are robust to objects deformations and partial occlusion. The object tracker is also able to detect collision and resolve it once the situation is over. Results of this investigation and a comparison between the proposed techniques and traditional methods are presented.