Browsing by Author "Gavrilova, Marina L."

Now showing 1 - 19 of 19
  • Thumbnail Image
    ItemOpen Access
    Adaptive mesh generation for terrain modeling and other applications
    (2004) Apu, Russel Ahmed; Gavrilova, Marina L.
    The thesis presents a real-time adaptive dynamic framework to model and render large terrain structures in real-time using the conventional 3D hardware. Terrain modeling is a highly regarded problem in the real-time rendering of three dimensional computer graphics. A terrain is a geometric representation of a landscape that has a large number of geometric primitives. Rendering terrain models using conventional algorithms is inefficient on a standard rendering hardware. The thesis provides data structures and corresponding algorithms for the efficient dynamic viewer dependent representation and rendering of large geometric models. The thesis utilizes the adaptive dynamic mesh model known as the Real-Time Optimally Adapting Mesh (ROAM) for terrain modeling and rendering. The main contribution of the thesis is a modified ROAM based model that offers better quality and higher efficiency by proposing a number of augments to the existing model. Efforts were made to improve the limitations of this model and extend the idea for the representation of larger data structures. Another major contribution of the thesis is the introduction of a novel adaptive dynamic mesh model called Adaptive Loop Subdivision. This new model is capable of rendering virtually infinitely large terrain meshes and offer an original smoothness control technique. Finally, the thesis provides efficient representation of the introduced models to solve real-world problems. A novel approach for motion planning based on the original dynamic mesh model called Adaptive Spatial Memory is presented. Experimental results and analysis confirm that the models and algorithms presented are effective and more efficient compared to other dynamic mesh models.
  • Thumbnail Image
    ItemOpen Access
    Balanced Multiresolution in Multilevel Focus+Context Visualization
    (2018-08-22) Hasan, Mahmudul; Samavati, Faramarz; Costa Sousa, Mário; Mudur, Sudhir Pandurang; Gavrilova, Marina L.; Jacob, Christian J.; Katz, Larry
    Given a set of symmetric/antisymmetric filter vectors containing only regular multiresolution filters, the method we present in this thesis can establish a balanced multiresolution (BMR) scheme for images, allowing their balanced decomposition and subsequent perfect reconstruction without the use of any extraordinary boundary filters. We define balanced multiresolution such that it allows balanced decomposition i.e. decomposition of a high-resolution image into a low-resolution image and corresponding details of equal size. Several applications of such a decomposition result in a balanced wavelet transform (BWT) that makes on-demand reconstruction of regions of interest (ROIs) efficient in both computational load and implementation aspects. We find such decomposition and perfect reconstruction based on an appropriate combination of symmetric/antisymmetric extensions near the image and detail boundaries. In our method, exploiting such extensions correlates to performing sample (pixel/voxel) split operations. We demonstrate our general approach for some commonly used symmetric/antisymmetric multiresolution filters. We also show the application of such a balanced multiresolution scheme in constructing an interactive multilevel focus+context visualization framework for the navigation and exploration of large-scale 2D and 3D images. Typically, the given filters are floating-point values, so our BWTs reversibly map integers to floating-point i.e. real values. We extend our balanced multiresolution framework further to construct reversible integer-to-integer BWTs from a given symmetric/antisymmetric decomposition filter vector of width less or equal to four. In our approach, we adjust the linear combination of fine samples suggested by the given decomposition vector using optimal sample split operations in combination with a rounding operation. Such adjustments translate an affine integer combination of fine samples to obtain an integer coarse sample, which closely approximates the floating-point coarse sample suggested by the given decomposition filter vector. The associated translation vectors give us the detail samples. Furthermore, when necessary, we construct every other detail sample differently in order to ensure local perfect reconstruction. Compared to their integer-to-real counterparts, the resulting reversible integer-to-integer BWTs occupy less memory, offer better compressibility, and do not require sample quantization for rendering purposes.
  • Thumbnail Image
    ItemOpen Access
    Cancelable Biometric System Based on Deep Learning
    (2020-09-24) Sudhakar, Tanuja; Gavrilova, Marina L.; Jacob, Christian P.; Henry, Ryan; Gavrilova, Marina L.
    With the increasing number of cyberattacks, Personal Identification Numbers (PINs), tokens, and passwords have been found to be insufficient for identity protection. Over the past decade, biometric systems have gained high popularity in providing secure mechanisms for user authentication. In this thesis, the safety of biometric data is rendered through the technique of ‘Cancelable Biometrics’. A cancelable biometric system for multi-instance biometrics has been designed with the use of deep learning. A deep learning architecture based on Convolutional Neural Network (CNN) and Multi Layer Perceptron (MLP) is presented to create a novel, accurate, and secure cancelable biometric solution. An implementation of the proposed solution has also been carried out on the cloud platform to provide a ubiquitous cancelable biometric authentication service. A high authentication accuracy, biometric template security and cancelability, fast response times, and cost efficiency are the merits of the presented cancelable biometric system.
  • Thumbnail Image
    ItemOpen Access
    Chaotic neural networks and multi-dimensional data analysis in biometric applications
    (2012) Ahmadian, Kushan; Gavrilova, Marina L.
    Humans have used body characteristics such as face, v01ce and gait. Although biometrics emerged from its extensive use in law enforcement (to identify criminals, to provide security clearance for employees in border protection, in fatherhood determination, forensics and positive identification of convicts), it is being increasingly used today to establish person recognition in a large number of civilian applications. In a practical biometric system (i.e., a system that employs biometrics for personal recognition), there are a number of important issues that should be considered, including performance (achievable recognition accuracy and speed) and circumvention (system resistance to noise and to being fooled by fraudulent methods). In order for biometric system to meet the above demands on performance and circumvention, more than one type of biometric is required. Hence, the need arises for the use of multi-modal biometrics, which is a combination of different biometric recognition technologies, varying from physical biometrics (such as face, iris and fingerprint recognition) to behavioral characteristics (i.e. signature, voice, and gate). Acquiring a group of different biometrics with different characteristics and specifications, results in a number of issues that should be addressed in a multi-modal biometric system. In such a system, one of the common problems is the high dimensionality of the data which impacts negatively system performance. Hence, dimensionality reduction methodologies are needed to be used. However, they have not been considered in recent multi-modal biometric systems due to gap between recently developed dimensionality reduction techniques in data mining and data analysis of biometric features. To remedy this situation, I propose a new methodology for shrinking down the finite search space of all possible subspaces by focusing on subspace analysis which is a novel approach in data clustering for biometric dataset. This is also a new contribution in biometric fusion methodology, which allows dealing with noisy data and makes the biometric system more error-proof. In summary, the purpose of this research is to develop a novel methodology based on the subspace clustering dimension reduction technique and chaotic neural network to improve the performance and circumvention of multi-modal biometric system. The focus is on the verification process where the proposed methodology is compared against some distinguished works on multi-modal biometrics. The system implementation and comparison criteria are included in this proposal to validate the developed multi-modal verification system.
  • Thumbnail Image
    ItemOpen Access
    Computational geometry methods in fingerprint identification
    (2006) Wang, Chengfeng; Gavrilova, Marina L.
  • Thumbnail Image
    ItemOpen Access
    Emerging Behavioral and Multi-Modal Biometric Approaches
    (2018-03-21) Rahman, Md Wasiur; Gavrilova, Marina L.; Alim, Usman; Hagen, Gregory; Alim, Usman R.
    Emerging behavioral biometrics play a vital role in situations where traditional biometrics may fail to identify the person correctly. This thesis focuses on designing new methods for biometric identification systems using emerging behavioral biometric characteristics such as EEG brain waves and Kinect gait. This research also develops multimodal biometric methods using the combination of emerging behavioral biometric characteristics with physiological biometric characteristics (such as face). The sub-goal of this research is to establish a relationship between a person's mental activity and a system identification accuracy. To fulfill this goal, this research studies overt and covert EEG signals. The second emerging behavioral biometric system is Kinect skeletal gait, where novelty of the work is in considering all possible joint-distance combinations. Finally, two multimodal systems are proposed and developed using the score-level fusion. The performance of the proposed identification systems is evaluated using five publicly available databases.
  • Thumbnail Image
    ItemOpen Access
    Facial model metamorphosis using geometrical methods
    (2006) Luo, Yuan; Gavrilova, Marina L.
  • Thumbnail Image
    ItemOpen Access
    Geometric approaches to path optimization and collision avoidance
    (2009) Hasan, Mahmudul; Gavrilova, Marina L.
  • Thumbnail Image
    ItemOpen Access
  • Thumbnail Image
    ItemOpen Access
    Kinect-based Gait Recognition Using Deep Learning
    (2020-11-27) Bari, A S M Hossain; Gavrilova, Marina L.; Jacobson, Michael J.; Korobenko, Artem
    Accurate gait recognition is of high significance for numerous industrial and consumer applications, including finance, virtual reality, online games, medical rehabilitation, collaborative space exploration, and others. This thesis proposes two novel deep learning architectures for a highly accurate and robust Kinect-based gait recognition. First, the architecture of the Deep Learning Neural Network (DLNN) is developed using two unique view and pose invariant geometric features. Second, the end-to-end training of a residual learning-based convolutional neural network, named KinectGaitNet, is proposed to enable even higher recognition performance without the necessity of extracting domain-specific handcrafted features. The performance of the DLNN architecture and KinectGaitNet are evaluated on two publicly available 3D skeleton-based gait datasets recorded with the Microsoft Kinect sensor. It is experimentally proven that the accuracy, precision, recall, and F-score of the DLNN architecture and KinectGaitNet are superior to the recent state-of-the-art methods for the Kinect skeleton-based gait recognition.
  • Thumbnail Image
    ItemOpen Access
    Multimodal Person Recognition using Social Behavioral Biometric
    (2018-04-06) Sultana, Madeena; Gavrilova, Marina L.; Yanushkevich, Svetlana N.; Leung, Henry; Li, Zongpeng; Ruhe, Guenther; Yang, Xiaosong
    The goal of a biometric recognition system is to make a human-like decision on individuals’ identity by recognizing their physiological and/or behavioral traits. Nevertheless, decision-making process by either a human or a biometric recognition system can be highly complicated due to low quality of data or an uncertain environment. Human brain has an advantage over computer system due to its ability to perform a massive parallel processing of auxiliary information such as visual cues, cognitive and social interactions, contextual and spatio-temporal data. Similarly to a human brain, social behavioral cues can aid the reliable decision-making of an automated biometric system. Being an integral part of human behavior, social interactions are likely to possess unique behavioral patterns. However, the significance of social behavior for automated user recognition has been noted in the scientific community only recently. In this doctoral thesis, a novel person recognition approach is presented that relies on the knowledge of individuals’ social behavior in order to enhance the performance of a traditional biometric system. The social behavioral information of individuals’ has been mined from an Online Social Network (OSN) and fused with traditional face and ear biometrics. This research identified a set of Social Behavioral Biometric (SBB) features from the online social information and proposed a framework to utilize these features for an automated person recognition for the first time. Extensive experiments confirm that human social behavior expressed through OSN can provide a unique insight onto person recognition. Performance of the proposed multimodal approach has been evaluated to determine the effectiveness of fusing social behavioral information. Experimental results on virtual and semi-real databases demonstrate significant performance gain in the proposed method over traditional biometric system. This doctoral research contributes to an emerging research direction in biometric domain as well as opens new frontier of studying social behavior in virtual domain.
  • Thumbnail Image
    ItemOpen Access
    Optimal path planning using spatial neighborhood properties
    (2007) Bhattacharya, Priyadarshi; Gavrilova, Marina L.
    The path planning problem, in its most well-know form, is the determination of a collision­free path for a mobile agent between source and destination. For practical path planning applications, a path should be as short as possible while maintaining a certain amount of clearance from obstacles. The problem is challenging as very few existing algorithms can fully satisfy both conditions. Very recently, improving path quality has received increased attention from researchers in the areas of GIS and robotics. In the first part of the thesis, we present a novel approach of combining path length with clearance from obstacles to generate an optimal path in just O(nlogn) time. In the second part, we provide a novel algorithm to solve the weighted region problem (wrp), concerned with determining an optimal path in a weighted terrain. Our algorithm is asymptotically faster than other existing approaches to solve the wrp and generates high quality optimal paths. We also provide a full description of the developed software system and extensive experimentation performed on GIS data.
  • Thumbnail Image
    ItemOpen Access
    Performance and integration of real time location system in healthcare
    (2011) Nayyar, Shikha; Gavrilova, Marina L.
  • Thumbnail Image
    ItemOpen Access
    Recognizing human emotional states from body movement
    (2019-07-09) Ahmed, Ferdous; Gavrilova, Marina L.; Korobenko, Artem; Rokne, Jon G.
    An emotion-aware computer system capable of responding to expressive human gestures and movements can significantly change the dynamics of human-computer interaction. This thesis addresses the problem of the creation of a computer model capable of automatically discerning emotion using various motion-related features of the human body. The proposed emotion recognition model automatically identifies relevant motion features using a combination of filter-based feature selection methods and the power of genetic algorithms. In addition to recognizing emotions, this thesis also focuses on gaining a deeper understanding of the role that various motion features play in emotion recognition, the ability to express emotionally relevant information by various parts of the human body and the effects of various action scenarios on emotion recognition. Rigorous analysis conducted on a proprietary dataset shows that the proposed computer model is very effective at identifying human emotion based predominantly on motion-related information. The proposed emotion recognition system also outperforms existing state-of-the-art computer models for emotion recognition.
  • Thumbnail Image
    ItemOpen Access
  • Thumbnail Image
    ItemOpen Access
    Spatial Partitioning for Distributed Path-Tracing Workloads
    (2018-09-21) Hornbeck, Haysn; Alim, Usman Raza; Gavrilova, Marina L.; Chan, Sonny
    The literature on path tracing has rarely explored distributing workload using distinct spatial partitions. This thesis corrects that by describing seven algorithms which use Voronoi cells to partition scene data. They were tested by simulating their performance with real-world data, and fitting the results to a model of how such partitions should behave. Analysis shows that image-centric partitioning outperforms other algorithms, with a few exceptions, and restricting Voronoi centroid movement leads to more efficient algorithms. The restricted algorithms also demonstrate excellent scaling properties. Potential refinements are discussed, such as voxelization and locality, but the tested algorithms are worth further exploration. The details of an implementation are outlined, as well.
  • Thumbnail Image
    ItemOpen Access
    Synthesizing techniques based on multiresolution
    (2007) Wecker, Lakin Christopher; Samavati, Faramarz F.; Gavrilova, Marina L.
  • Thumbnail Image
    ItemOpen Access
    Theories and Experiments of Cognitive Knowledge Bases for Machine Learning
    (2018-06-26) Zatarain Duran, Omar Ali; Wang, Yingxu; Gavrilova, Marina L.; Fapojuwo, Abraham Olatunji; Chen, Zhangxing; Budin, Gerhard
    This thesis presents a framework of studies on theories, methodologies, algorithms, and experiments on cognitive knowledge bases (CKBs) for machine knowledge learning in cognitive computing and computational linguistics. CKB is both the results and the means of machine learning methodologies mimicking human learning and semantic comprehensions. Technologies for machine learning can be classified into six categories according to Dr. Y. Wang known as object identification, cluster classification, pattern recognition, functional regression, behavior generation, and knowledge acquisition. Most current machine learning techniques fall into the first five categories. However, the sixth category of knowledge learning as humans do has remained as a fundamental problem and challenge in machine learning, AI, and computational intelligence. A set of algorithms, tools, and experiments on machine knowledge learning is designed in order to demonstrate that cognitive machines may create their own concepts and CKBs through knowledge learning. The accuracy and cohesiveness of machine learnt results may outperform humans. This leads to the implementation of formal knowledge comprehension and quantitative semantic analyses by cognitive systems based on CKBs and machine semantic comprehensions. The theoretical framework and case studies derived from this research will impact the field of machine knowledge learning technologies and the development of novel cognitive systems. This research will enable industrial applications such as personal leaning assistants, cognitive search engines, and cognitive translators.
  • Thumbnail Image
    ItemOpen Access
    Theories and Methodologies for Cognitive Machine Learning based on Denotational Mathematics
    (2018-06-22) Valipour, Mehrdad; Wang, Yingxu; Gavrilova, Marina L.; Yanushkevich, Svetlana N.; Chen, Zhangxing; Chen, Liang
    Learning is a cognitive process of knowledge and behavior acquisition for both humans and machines. Cognitive machine learning systems are increasingly demanded in modern knowledge-based industry, society, and everyday lives. This study on theories and applications of cognitive machine learning based on denotational mathematics is designed to develop methodologies, algorithms, and their implementations for machine enabled knowledge learning at the conceptual, phrasal, and sentence levels via cognitive computing technologies. The main objectives of this work are: a) To develop a cognitive and mathematics-based machine learning theory for knowledge acquisition and semantic manipulations; b) To enable machines to autonomously learn and understand semantics expressed in natural languages underpinned by unsupervised cognitive computing algorithms; and c) To design and implement a brain-inspired cognitive learning engine for inductively learning from the level of formal concepts to those of phrases and sentences. The thesis is embodied by three novel and autonomous machine knowledge learning algorithms underpinned by Wang’s denotational mathematics. In this research, properties of formal concepts and mathematical rules of concept algebra are formally studied. A method for building quantitative semantic hierarchies of formal concepts is implemented by cognitive machine learning. Theories and mathematical models for an unsupervised algorithm of phrase learning are developed based on rigorous concept comprehensions by cognitive machine learning. A machine knowledge learning system for sentence syntactic analysis and semantic synthesis is developed and implemented by novel cognitive computing technologies. This thesis does not only present a set of basic studies on machine learning challenges in the sixth category of knowledge learning and semantic comprehension, but also implement efficient cognitive machine learning systems mimicking human learning. This research will enable a wide range of industrial applications such as cognitive robotics, natural language comprehension systems, personal leaning assistants, cognitive search engines, and language translators.