Doctor of Philosophy, Case Western Reserve University, 2006, Electrical Engineering

I present a new method of visual search for objects that include straight lines. This is usually the case for machine-made objects. I describe existing machine vision search methods and show how my method of visual search gives better performance on objects that have straight lines. Inspired from human vision, a two-step process is used. First, straight line segments are detected in an image and characterized by their length, mid-point location, and orientation. Second, hypotheses that a particular straight line segment belongs to a known object are generated and tested. The set of hypotheses is constrained by spatial relationships in the known objects. I discuss implementation of my method and its performance and limitations in real and synthetic images. The speed and robustness of my method make it immediately applicable to many machine vision problems.

Committee: Christos Papachristou (Advisor) Subjects: Computer Science

PHD, Kent State University, 2024, College of Arts and Sciences / Department of Computer Science

With the increasing adoption of multi-camera setups for comprehensive monitoring, such as trauma rooms in hospitals, privacy leakage in video surveillance systems has become a significant concern.This research aims to develop an intelligent video surveillance system that leverages skeletal based methods to ensure privacy protection while enabling accurate action recognition. I address the challenge of privacy protection by detecting and blurring human heads or replacing individuals in the system with skeleton representations, while simultaneously applying weighted fusion methods to enhance the action recognition.The main challenge in existing skeletal-based action recognition algorithms is about the insufficient accuracy. To address this issue, I propose two novel algorithms to fuse the results from multiple cameras. The first idea is to divide the floor into many grids, then give them with different fusion weights according to the performance of skeletons in different grids. The second method involves integrating with the YOLO system to intelligently recognize the orientation of individuals within the camera's field of view, then different weights are assigned to the same person when they are captured from different camera perspectives.Those fusion approaches seeks to obtain more reliable and precise action recognition results. By fusing data from various angles, the system enhances the robustness of action recognition, making it suitable for real-world applications. The research also involves extensive experimentation and data analysis to evaluate the proposed algorithm's performance and compare it with existing methods. I aim to achieve a significant improvement in the accuracy of action recognition while ensuring the protection of individuals' privacy in the surveillance context.

Committee: Qiang Guan (Advisor); Qiang Guan (Committee Chair); Sara Bayramzadeh (Committee Member); Ruoming Jin (Committee Member); Lei Xu (Committee Member); Kambiz Ghazinour (Committee Member) Subjects: Computer Science

Master of Science, The Ohio State University, 2023, Mechanical Engineering

Soft grippers have been a major area of research in the field of robotics over the past years. Soft grippers have the advantage of being able to grasp a multitude of different objects whether fragile or stiff compared to their counterparts of rigid grippers due to their flexibility and softer material allowing them to conform to the objects. However, drawbacks of a soft gripper are its precision and payload when picking up an object. In this research, a three-fingered pneumatic soft gripper is constructed with a 3D printer and tested experimentally for its grasp quality on a certain object. This gripper utilizes its geometry to create a bending moment when actuated to change shape and grasp force based on the amount of air pressure provided. Methodology was created to integrate data collection between the pose of the object and the pose of the gripper. Two tests were conducted for different pressure configurations of the gripper and different weights of the object. The first test consisted of an accuracy test to determine if the object's position was at its expected location. The second test consisted of a stability test where the grasped object underwent an acceleration test. A pressure configuration needed to pass both tests to qualify as a successful grasp. With zero weight added to the object, there were a total of six pressure configurations with a successful grasp after the 1m/s2 stability test. There were three pressure configurations each for 50g and 100g added to the object when looking at the 1m/s2 stability test. However, there were zero successful grasp configurations between the accuracy and 4m/s2 stability tests. This research provides a data collection process and experiment methodology that can be used to quantify the quality of a grasp in an experimental setting.

Committee: Haijun Su Ph.D (Advisor); Ayonga Hereid Ph.D (Committee Member) Subjects: Mechanical Engineering; Robotics

Doctor of Philosophy (PhD), Wright State University, 2022, Computer Science and Engineering PhD

Activities of Daily Living (ADL's) are the activities that people perform every day in their home as part of their typical routine. The in-home, automated monitoring of ADL's has broad utility for intelligent systems that enable independent living for the elderly and mentally or physically disabled individuals. With rising interest in electronic health (e-Health) and mobile health (m-Health) technology, opportunities abound for the integration of activity monitoring systems into these newer forms of healthcare. In this dissertation we propose a novel system for describing 's based on video collected from a wearable camera. Most in-home activities are naturally defined by interaction with objects. We leverage these object-centric activity definitions to develop a set of rules for a Fuzzy Inference System (FIS) that uses video features and the identification of objects to identify and classify activities. Further, we demonstrate that the use of FIS enhances the reliability of the system and provides enhanced explainability and interpretability of results over popular machine-learning classifiers due to the linguistic nature of fuzzy systems.

Committee: Tanvi Banerjee Ph.D. (Advisor); Yong Pei Ph.D. (Committee Member); Michael Riley Ph.D. (Committee Member); Mateen Rizki Ph.D. (Committee Member); Thomas Wischgoll Ph.D. (Committee Member) Subjects: Computer Science

PhD, University of Cincinnati, 2020, Engineering and Applied Science: Electrical Engineering

Action recognition is an important computer vision task. It focuses on identifying the behavior or the action performed by humans from images. Action recognition using various wearable sensors and videos is a well studied and well established topic. This thesis focuses on action recognition in still images, a new and challenging area of research. For example, understanding motion from static images is a difficult task as spatio-temporal features that is most commonly used for predicting actions is not available. Action recognition in still images has a variety of applications such as searching for frames in videos using action, searching a database of images using an action label, surveillance, robotic applications etc. It can also be used to give a more meaningful description of the image. The goal of this thesis is to perform action recognition in still images and infer object affordances by characterizing the interaction between the human and the object. Object affordance refers to determining the use of an object based on its physical properties. The main idea is to learn high level concepts such as action and object affordance by extracting information of the objects and their interactions in an image.

Committee: Anca Ralescu Ph.D. (Committee Chair); Kenneth Berman Ph.D. (Committee Member); Rashmi Jha Ph.D. (Committee Member); Wen-Ben Jone Ph.D. (Committee Member); Dan Ralescu Ph.D. (Committee Member) Subjects: Electrical Engineering

Master of Science (MS), Wright State University, 2019, Human Factors and Industrial/Organizational Psychology MS

Object recognition entails a complex interplay between top-down and bottom-up signals. Yet, limited research has investigated the mechanisms through which top-down processes, such as task context and behavioral goals impact the neural basis of visual object processing. Using electroencephalography (EEG), we studied the temporal dynamics of task and object processing to identify how early the impact of task can be observed. We recorded ERPs from participants as they viewed object images from four categories spanning animacy (Inanimate: roller-skate, motorbike; Animate: cow, butterfly) and size (Large: motorbike, cow; Small: roller-skate, butterfly) dimensions under four task conditions comprising conceptual (naturalness, size) and perceptual (color, tilt) dimensions. We did not find evidence of behavioral goals, as manipulated by the task context, modulating early visual object representations, as indexed by early visual ERPs (P1, N1, P2), in extrastriate cortex. Additional analyses revealed that task-related processing occurred predominately in later time windows (300-600ms) within frontoparietal regions. Irrespective of task, we also observed a variety of object category effects across early visual ERPs. These findings support previous neuroimaging results suggesting object representations in occipitotemporal cortex are organized based on their animacy and real-world size, and, importantly, these ERP results indicate these organizational principles can be observed in relatively early stages along the visual processing hierarchy. Taken together, this work adds to the body of psychological and neuroscientific research examining how and when top-down and bottom-up signals interact to form the basis of visual object processing, facilitating of high-level vision.

Committee: Assaf Harel Ph.D. (Advisor); Joseph W. Houpt Ph.D. (Committee Member); Ion Juvina Ph.D. (Committee Member) Subjects: Cognitive Psychology; Neurosciences; Psychology

Doctor of Philosophy (PhD), Ohio University, 2019, Electrical Engineering & Computer Science (Engineering and Technology)

Artificial intelligence (AI) is being widely applied to various practical problems, and researchers are working to address numerous issues facing the field. The organizational structure and learning mechanism of the memory is one such issue. A cognitive agent builds a representation of its environment and remembers its experiences to interpret its inputs and implements its goals through its actions. By doing so it demonstrates its intelligence (if any), and it is its learning mechanism, value system and sensory motor coordination that makes all this possible. Memory in a cognitive agent stores its knowledge, knowledge gained over a life-time of experiences in a specific environment. That is, memory includes the “facts”, the relationships between them, and the mechanism used to learn, recognize, and recall based on the agent's interaction with the world/environment. It remembers events that the agent experienced reflecting important actions and observations. It motivates the agent to do anything by providing assessment of the state of the environment and its own state. It allows it to plan and anticipate. And finally, it allows the agent to reflect on itself as an independent being. Hence, memory is critical for intelligence, for it is the memory that determines a cognitive agent's abilities and learning skills. Research has shown that while memory in humans can be classified into different types, based on factors such as their longevity and cognitive mechanisms used to create and retrieve them, they all are achieved using a similar underlying structure. The focus of this dissertation was on using this principle, i.e. different memories created using the same underlying structure, to implement memory for cognitive agents using a biologically plausible model of neuron. This work was an attempt to demonstrate the feasibility of implementing self-organizing memory structures capable of performing the various memory related tasks necessary for a cognitive agent using a c (open full item for complete abstract)

Committee: Wojciech Jadwisienczak (Advisor) Subjects: Electrical Engineering

Master of Sciences, Case Western Reserve University, 2019, EECS - Computer and Information Sciences

Computer vision tasks require collecting large volumes of data, which can be a time consuming effort. Automating the collection process with simulations speeds up the process, at the cost of the virtual data not closely matching the physical data. Building upon a previous attempt to bridge this gap, this thesis proposes three nuances to improve the correspondence between simulated and physical 3-D point clouds and depth images. First, the same CAD files used for simulated data acquisition are also used to 3-D print physical models used for physical data acquisition. Second, a new projection method is developed to make better use of all information provided by the depth camera. Finally, all projection parameters are unified to prevent the deep learning model from developing a dependence on intensity scaling. A convolutional neural network is trained on the simulated data and evaluated on the physical data to determine the model's generalization ability.

Committee: Wyatt Newman Dr. (Advisor); Michael Lewicki Dr. (Committee Member); Cenk Cavusoglu Dr. (Committee Member) Subjects: Computer Science; Robotics

Doctor of Philosophy, The Ohio State University, 2018, Electrical and Computer Engineering

Video understanding is able to accomplish various tasks which are fundamental to human-robot interaction and detection. Such tasks include object tracking, action recognition, object detection, and segmentation. However, due to the large data volume in video sequence and the high complexity of visual algorithms, most visual algorithms suffer from low robustness to maintain a high efficiency, especially when it comes to the real-time application. It is challenging to achieve high robustness with high efficiency for video understanding. In this dissertation, we explore the efficient and robust video understanding for human-robot interaction and detection. Two important applications are the health-risky behavior detection and human tracking for human following robots. As large portions of world population are approaching old age, an increasing number of healthcare issues arise from unsafe abnormal behaviors such as falling and staggering. A system that can detect the health-risky abnormal behavior of the elderly is thus of significant importance. In order to detect the abnormal behvior with high accuracy and timely response, visual action recognition is explored and integrated with inertial sensor based behavior detection. The inertial sensor based behavior detection is integrated with a visual behavior detection algorithm to not only choose a small volume of the video sequence but also provide a likelihood guide for different behaviors. The system works in a trigger-verification manner. An elder-carried mobile devices either by a dedicated design or a smartphone, equipped with inertial sensor is used to trigger the selection of relevant video data. The selected data is then fed into visual verification module, and in this way the selective utilization of video data is achieved and the efficiency is guaranteed. By using selected data, the system is allowed to perform more complex visual analysis and achieve a higher accuracy. A novel tracking approach for robus (open full item for complete abstract)

Committee: Yuan Zheng (Advisor); Dong Xuan (Committee Member); Yuejie Chi (Committee Member) Subjects: Computer Engineering; Computer Science

Master of Sciences, Case Western Reserve University, 2017, EMC - Mechanical Engineering

This thesis focuses on deep-learning approaches to recognition and pose estimation of graspable objects using depth information. Recognition and orientation detection from depth-only data is encoded by a carefully designed 2D descriptor from 3D point clouds. Deep-learning approaches are explored from two main directions: supervised learning and semi-supervised learning. The disadvantages of supervised learning approaches drive the exploration of unsupervised pretraining. By learning good representations embedded in early layers, subsequent layers can be trained faster and with better performance. An understanding of learning processes from a probabilistic perspective is concluded, and it paves the way for developing networks based on Bayesian models, including Variational Auto-Encoders. Exploitation of knowledge transfer--re-using parameters learned from alternative training data--is shown to be effective in the present application.

Committee: Wyatt Newman PhD (Advisor); M. Cenk Çavusoglu PhD (Committee Member); Roger Quinn PhD (Committee Member) Subjects: Computer Science; Medical Imaging; Nanoscience; Robotics

Doctor of Philosophy, The Ohio State University, 2017, Computer Science and Engineering

Visual infrastructure, which consists of connected visual sensors, has been extensively deployed and is vital for various important applications, such as surveillance, tracking, and monitoring. However, there are still many problems regarding visual sensor deployment for optimal coverage and visual data processing technology. Challenges remain with the sectoral visual sensing model, the complexity of image processing, and these sensors' vulnerability to noisy environments. Solving these problems will improve the performance of visual infrastructure, which increases accuracy and efficiency for these applications. This dissertation focuses on visual-infrastructure-related technologies. In particular, we study the following problems. First, we study visual infrastructure deployment. We propose local face-view barrier coverage (L-Faceview), a novel concept that achieves statistical barrier coverage in visual sensor networks leveraging mobile objects' trajectory information. We derive a rigorous probability bound for this coverage via a feasible deployment pattern. The proposed detection probability bound and deployment pattern can guide practical camera sensor deployments in visual infrastructure with limited budgets. Second, we study visual-infrastructure-based object recognition. We design and implement R-Focus, a platform with visual sensors that detects and verifies a person holding a mobile phone nearby with the assistance of electronic sensors. R-Focus performs visual and electronic data collection and rotates based on the collected data. It uses the electronic identity information to gather visual identity information. R-Focus can serve as a component of visual infrastructure that performs object identity recognition. Third, we study visual-infrastructure-based object localization. We design Flash-Loc, an accurate indoor localization system leveraging flashes of light to localize objects in areas with deployed visual infrastructure. An object emits a seq (open full item for complete abstract)

Committee: Dong Xuan (Advisor); Yuanfang Zheng (Committee Member); Ten-Hwang Lai (Committee Member) Subjects: Computer Engineering; Computer Science

Doctor of Philosophy (Ph.D.), University of Dayton, 2014, Electrical Engineering

Object detection and tracking in imagery captured by aerial systems are becoming increasingly important in computer vision research. In aerial imagery, objects can appear in any orientation, varying sizes and in different lighting conditions. Due to the exponential growth in sensor technology, increasing size and resolution of aerial imagery are becoming a challenge for real-time computation. A rotation invariant feature extraction technique for detecting and tracking objects in aerial imagery is presented in this dissertation. Rotation invariance in the feature representation is addressed by considering concentric circular regions centered at visually salient locations of the object. The intensity distribution characteristics of the object region are used to represent an object effectively. A set of intensity-based features is derived from intensity histograms of the circular regions and they are inherently rotation invariant. An integral histogram computation approach is used to compute these features efficiently. To improve the representational strength of the feature set for rotation and illumination-invariant object detection, a gradient-based feature set is derived from normalized gradient orientation histograms of concentric regions. Rotation invariance is achieved by considering the magnitude of the Discrete Fourier Transform (DFT) of the gradient orientation histograms. A novel computational framework called Integral DFT is presented for fast and efficient extraction of gradient-based features in large imagery. A part-based model, which relies on a representation of an object as an aggregation of significant parts, using the gradient-based features is also presented in this dissertation. Integrating the features of significant parts gives robustness to partial occlusions and slight deformations, thus leading to a better object representation. The effectiveness of the intensity-based feature is demonstrated in tracking objects in Wide Area Motion Imagery (WA (open full item for complete abstract)

Committee: Vijayan Asari (Committee Chair); Keigo Hirakawa (Committee Member); Raul Ordonez (Committee Member); Youssef Raffoul (Committee Member) Subjects: Computer Engineering; Computer Science; Electrical Engineering; Engineering

Master of Arts, The Ohio State University, 2013, Psychology

The different types of non-accidental properties have been commonly treated as having equal perceptual salience. In an effort to test this hypothesis, discrimination thresholds were measured for four different types of transformations that altered the Euclidean, affine, projective, or topological structure of two basic 2D objects. The results confirm that there are significant differences in the relative perceptual salience of these transformations such that the topological change was the most salient followed by projective change which was followed by affine change. Changes in Euclidean structure were the most difficult to detect. To determine how well existing metrics account for the observer data, these thresholds were compared to the predictions from pixel distance, pixel angle, correlation, total area, and Hausdorff metrics. The results suggest that all the tested metrics are poor predictors of observer data for the perception of non-accidental properties.

Committee: James Todd (Advisor) Subjects: Cognitive Psychology

PhD, University of Cincinnati, 2003, Engineering : Electrical Engineering

Image segmentation is the premise of image understanding process. Deformable contour methods (DCMs) are currently the most popular image segmentation approaches and many of them were proposed in past years. In order to understand the strengths and weakness of different DCMs on image segmentation applications, this dissertation provides a qualitative and quantitative comparison of some major DCMs on a set of selected biomedical images. Though they works well on some cases, there are still many very challenging and difficult problems that they cant handle, such as very blur contour segments, complex shape, inhomogeneous interior and inhomogeneous contour region distribution, just to name a few. Model based DCMs are necessary to solve these problems. We present a new model based approach for accurate shape recovery from images by applying a skeleton based shape matching method. The shape matching method consists of two major operations: skeleton extraction and shape model representation, and matching and model detection. For skeleton extraction, a distance transformation based method is employed. The shape model of an object consists of both the skeleton model and the contour segments model, which are used in tandem and in a complementary manner. The skeleton matching algorithm is introduced to match the skeleton of a DCM contour against a set of object skeleton models to select the candidate model and determine the corresponding landmarks on the contours based on their skeleton structure and a similarity function. In shape recovery process, segments obtained from these landmarks are then matched against the detected model segments for errors. For any large error in segments mismatch, a fine-tuning process, which is formulated as a maximization of a posteriori probability, given the contour segments model and image features, is performed for final result. The skeleton based shape matching approach is also amendable for object recognition. The skeleton matching algorith (open full item for complete abstract)

Committee: Dr. William G. Wee (Advisor) Subjects:

Master of Arts, The Ohio State University, 2011, Psychology

One common method of representing images is to reduce an image to a collection of features. Many simple features have been proposed, such as pixel intensities and wavelet responses, but these choices are fundamentally unsuitable for capturing the configural relations of objects and object parts, as spatial information associated with each feature is lost. Another recent strategy, known as "feature-hierarchy" modeling, involves the use of overlapping, redundant features. These features are obtained by processing an image across a hierarchy of units tuned to progressively more complex properties. An open question is whether such approaches produce data structures rich enough for implicitly capturing configural relations. We implemented three experiments and several computer simulations to address this issue. Our method involved the use of four classes of objects, each derived from the simple spatial relationships present in classic Vernier and bisection acuity tasks. All human observers achieved near perfect categorization performance after relatively few exposures to each stimulus class. This ability also transferred across several dimensions, including orientation and background context. By contrast, simulations on a feature-hierarchy model revealed poor performance for this class of models. Furthermore, the moderate categorization accuracy achieved did not transfer across even the simplest of dimensions. These results indicate that this approach to image representation lacks a fundamental property necessary for encoding the spatial configurations of object parts.

Committee: Alexander Petrov (Advisor); James Todd (Committee Member); Dirk Bernhardt-Walther (Committee Member) Subjects: Behavioral Psychology; Behavioral Sciences; Cognitive Psychology

Doctor of Philosophy, The Ohio State University, 2008, Geodetic Science and Surveying

Object recognition is one of the key processes in Photogrammetry to generate maps from sensory information, because it is to convert 'data' to 'information.' However, as the size of input data is increased, it also has been one of the bottle neck processes in Photogrammetry. Thus many researchers have been working on developing semi-automated or automated object recognition methods to speed up the process. Some of the developed methods have been proved to be feasible in controlled environments, and others have been applicable for real world applications. However, most of the conventional object recognition methods still require human operators' interventions to correct errors or clarify ambiguous results.The new object recognition method proposed in this dissertation is to recognize multiple types of objects in parallel so that the ambiguous results would be minimized. Since 1980's, new paradigms in Computer Science such as parallel computing and agent models have emerged with the progress of computer systems. The proposed method is built on one of the paradigms, the agent model. With built-in knowledge and predefined goals, the agent actively searches clues to reach the goals. In a multi-agent model, several agents with specific goals and tasks are deployed, and they are trying to reach the main goal together. The proposed system consists of the coordinator agent, the recognition agents, and the data agent. The coordinator agent initiates other agents, and the data agent handles and processes input data. While the recognition agents aggressively collect regions for the target objects, sometimes conflicts arise between more than two recognition agents. With the proposed conflict resolution scheme, the conflicts can be resolved, and finally ambiguity can be removed. Experiments on the proposed system were performed with a multi-spectral image and LIDAR data. Results of feature extraction done by the data agent, and object recognition are presented. The results show t (open full item for complete abstract)

Committee: Anton Schenk PhD (Advisor); Alan Saalfeld PhD (Committee Member); Alper Yilmaz PhD (Committee Member); Bea Csatho PhD (Committee Member) Subjects: Computer Science

Master of Science (MS), Ohio University, 1991, Electrical Engineering & Computer Science (Engineering and Technology)

Fast implementation of hadamard transform for object recognition and classification using parallel processor

Committee: M. Celenk (Advisor) Subjects:

Master of Science (MS), Ohio University, 1989, Electrical Engineering & Computer Science (Engineering and Technology)

Hypercube machine implementation of a 2-D FFT algorithm for object recognition

Committee: Mehmet Celenk (Advisor) Subjects:

Doctor of Philosophy (PhD), Ohio University, 2004, Electrical Engineering & Computer Science (Engineering and Technology)

Landmark recognition, identified as one of the most important research areas in robot navigation, has heretofore mainly focused on simple landmarks, limiting the use of robots in complex environments. In this dissertation, the problem of generalized landmark recognition in complex environments is addressed. A complete framework is presented for landmark recognition and scene understanding with a combination of data-driven and model-driven approaches. In the data-driven approach, natural scene analysis is performed using a proposed texture model called intensity interactive maps (IIM). Natural scene synthesis is then studied via multi-eigenspace decomposition that is useful in differentiating background from objects. The model-driven approach achieves recognition by removing the background while processing only potential objects. Two approaches are presented for object detection and recognition: a multilevel Markov Random Field (MRF)based model and an active contour model that integrates color, texture and shape priors. To address dynamic environment changes and illumination variances, a feedback strategy is introduced and modeled into both the multilevel MRF model and the active contour model to achieve illumination adaptation. Various experiments illustrate the performance of the proposed framework. The segmentation results in natural scene analysis are a major step forward, unlike previous results, they include semi-meaningful segments consistent with human perception. The results of natural scene synthesis are also perceptually acceptable, which has not been achieved by other texture synthesis based technology. The multi-level MRF model demonstrates robust performance for object detection and recognition under changing environments. The active contour model with prior knowledge can reliably detect and segment an object in a complex background, a cluttered environment, and a scene with partial occlusion. Although the framework was initially proposed to address com (open full item for complete abstract)

Committee: David Chelberg (Advisor) Subjects:

Master of Engineering, Case Western Reserve University, 2011, EMC - Mechanical Engineering

This thesis presents a novel object recognition engine for the application of bin-picking. The algorithm is capable of quickly recognizing and estimating the pose of objects in a given unorganized scene. Based on the oriented point-pair feature vector, the algorithm matches points in the scene to points on the surface of an original model via an efficient voting process. Groups of features defining a point in the scene are used to find probable matching model points in a precompiled database. Sets of candidate model and scene point-pair matches are created and then filtered based on a geometric consistency constraint. Results show that the algorithm can produce centroid error values of less than ≈.55mm and angular error values of less than ≈4° without a secondary iterative closest point algorithm. Run-times are in the range of .1 to .5 secs to locate a single object.

Committee: Roger Quinn PhD (Advisor); Frank Merat PhD (Committee Member); Jaikrishnan Kadambi PhD (Committee Member) Subjects: Robotics