Computer vision (CV) methods have grown significantly in recent years due to the rapid development of artificial intelligence and machine learning. They involve the ability of machines to interpret, analyze, and understand visual data from the environment, including images, videos, and 3D data captured by cameras, sensors, and other devices. CV algorithms lend themselves as effective tools in real-world applications, e.g., healthcare, robotics, and infrastructure resilience. In this dissertation, we focus on their ability to detect anomalies in different domains including predictive analysis for healthcare, anomalous activity recognition in videos, and built environment monitoring. Traditionally, CV algorithms have been deployed in a centralized fashion where data is aggregated in a central server and the training is conducted centrally. Although centralized CV has shown significant potential for solving numerous challenging tasks, it faces some challenges, e.g., privacy concerns, data sharing limitations, bandwidth constraints, and computational complexity. Moreover, the prevalent integration of edge devices increases the complexity of the centralized CV problems. Therefore, with the advent of distributed learning frameworks, such as federated learning, there has been a paradigm shift toward decentralized computation. Distributed computation addresses the challenges of traditional centralized CV and enhances scalability, privacy, and real-time adaptability. Additionally, the recent edge devices have become computationally sufficient to carry out small computational tasks and research on distributed or edge computation has become a necessary demand. This dissertation initially reveals the development of novel computer vision algorithms and their applications for solving different real-world problems, such as COVID-19 infection detection from chest radiographic images, house detection from aerial imagery using Faster Region-based Convolutional Neural Network (Faster-RCNN), and unsupervised anomaly detection in videos. To deal with the mentioned challenges of centralized CV algorithms, the second part of this dissertation focuses on distributed CV solutions. To this end, we propose to deploy distributed learning algorithms to enable agent-based anomaly detection and the development of novel distributed learning algorithms. It includes the federated transfer learning (FTL) framework for the optimization of resource-constraints heterogeneous devices. Further, it develops a federated multimodal microscopic traffic simulation (FedMMTS) framework that incorporates federated learning to preserve data privacy and applies multimodal data analytics to achieve more accurate sensing of the environment in a microscopic traffic simulation scenario. Last but not the least, agent-based anomaly detection in videos is explored in this dissertation.

Khandaker Mamun Ahmed is a Ph.D. candidate in Sustainability, Optimization, and Learning for InterDependent networks laboratory (solid lab), Knight Foundation School of Computing and Information Sciences (KFSCIS) at Florida International University (FIU) under the supervision of Dr. M. Hadi Amini. His research interests span Computer Vision, Federated Learning, Internet of Things, Machine learning, and their applications in anomaly detection. Khandaker received his B.Sc. in Software Engineering from the University of Dhaka, Bangladesh, and received his Master’s in Computer Science from FIU. He received the “2022 Best Graduate Student Research Award” from KFSCIS and three travel scholarships to present his research. Khandaker is a co-inventor of a patent and published multiple peer-reviewed articles in selective venues, including ICMLA, BIBM, and ICCCN. He also mentored several graduate and undergraduate students, delivered a workshop on computer vision for suspicious activity recognition, and was actively involved in NSF RET programs.