Stelios Krinidis

Urban Internet-of-Things (IoT) infrastructures increasingly host distributed artificial intelligence workloads across heterogeneous edge environments, supporting applications such as urban monitoring, resource optimization, and large-scale sensing analytics. Despite the rapid adoption of edge AI, current systems lack a unified architectural layer responsible for orchestrating distributed intelligence across dynamic urban infrastructures. Existing orchestration mechanisms are typically rule-based and rely on low-level telemetry signals such as latency, node availability, or network conditions, limiting their ability to interpret the contextual meaning of system behavior and environmental events. This paper introduces Urban AI OS, a trustworthy LLM-based semantic control layer for distributed urban edge intelligence. The proposed architecture transforms heterogeneous telemetry and environmental signals into high-level semantic events through large language model reasoning, enabling context-aware orchestration decisions including adaptive topology management, workload coordination, and node role assignment. To address the reliability risks associated with LLM-assisted control, Urban AI OS employs confidence-gated execution with deterministic fallback policies, post-action monitoring, rollback, and auditable decision logging to constrain the operational impact of uncertain or erroneous LLM recommendations. By decoupling semantic reasoning from the data plane execution of AI workloads, Urban AI OS provides a model-agnostic framework for managing large-scale urban edge intelligence infrastructures. Experimental evaluation on a real-world urban IoT deployment with 50+ edge devices demonstrates that the proposed semantic operating layer improves system responsiveness by 35%, reduces unnecessary topology changes by 42%, and improves communication efficiency compared to conventional rule-based and adaptive threshold baselines, while maintaining confidence calibration through formal fallback mechanisms.

Resource-constrained edge devices require efficient and adaptive health monitoring to ensure reliable operation under dynamic workloads. This paper proposes an LLM-driven, context-aware framework that integrates real-time system metrics, historical data, and task-specific importance levels for anomaly detection and prediction. Specifically, the framework forecasts the semantic health state of the edge device five minutes ahead based on recent monitoring observations, enabling proactive intervention before critical operating conditions are reached. Multidimensional metrics, including CPU, memory, temperature, network load, and process information, are transformed into structured time-series representations and used as input to pre-trained machine learning models. A large language model (LLM) acts as an orchestration layer, dynamically selecting the most appropriate predictive model based on system context and resource constraints. This enables adaptive trade-offs between accuracy, latency, and computational cost. Experimental results on Raspberry Pi devices show that the proposed approach achieves comparable or improved performance while reducing resource usage compared to static methods.

Computer vision and sensor technology advancements have enabled significant progress in 3D reconstruction. However, critical challenges remain when extending these methods to underwater environments. This paper addresses the disparities between terrestrial and underwater 3D reconstruction by presenting a novel comparative framework that evaluates identical objects in these distinct settings under real-world conditions. To mitigate underwater-specific challenges such as light absorption, scattering, and reduced image clarity, a new enhancement methodology combining RGB stretching and Contrast Limited Adaptive Histogram Equalization (CLAHE) is proposed. Experimental results demonstrate that the enhancement increased processing time by 2.99% compared to the underwater model, while improving reconstructed points by 5.05%, detected features by 3.12%, and reconstructed features by 6.58%. Furthermore, structural metrics such as smoothness, genus connectivity, and average skewness showed significant positive improvements. Furthermore, structural metrics showed significant improvements with genus connectivity increasing by 1.29%, smoothness improving by 0.69%, and average skewness increasing by 0.18%, based on the raw underwater model and underwater enhanced model. By providing a direct cross-environment comparison, this study bridges the gap between terrestrial and underwater 3D reconstruction, offering critical insights into methodological adaptations necessary for robust and reliable modeling in diverse environmental conditions. The proposed framework lays the foundation for advancing underwater 3D reconstruction technology to achieve parity with its terrestrial counterpart, with applications spanning cultural heritage preservation, marine exploration, and environmental monitoring.

Underwater images often suffer from degradations, including color distortion, reduced visibility, and low contrast due to light absorption and scatter in water. Numerous enhancement techniques have been proposed to improve visual quality and address these challenges. However, no single method consistently performs best across all underwater scenes. This work introduces a novel deep learning framework for the automatic selection of the most suitable enhancement technique for underwater images. A novel fused objective metric, combining the Underwater Color Image Quality Evaluation (UCIQE), Underwater Image Quality Measure (UIQM), and Underwater Image Fidelity (UIF) metrics is introduced to assess image quality effectively. The metric is then utilized to train a Shifted Window (Swin) transformer model, which predicts the best enhancement method for each image. This approach advances automatic underwater image enhancement by addressing varying image conditions with a data-driven, adaptive process. Experimental results show that the proposed model achieves an F1 score of 87.88% in selecting the optimal enhancement technique, effectively determining the best enhancement based on the characteristics of the input image.

Digital Twins for cultural heritage buildings commonly depend on high-fidelity 3D scanning, detailed onsite surveys, and unrestricted data acquisition. In many countries, however, legal, regulatory, and conservation constraints render such methods inaccessible or explicitly prohibited, significantly limiting the deployment of digital-heritage technologies in real settings. This paper introduces HeritageTwin Lite, a regulation-compliant workflow for constructing low-detail yet operational Digital Twins of protected cultural heritage buildings using only publicly permissible data sources. The proposed approach relies on a GIS-based 2D application through which users select a site and manually delineate building footprints and structural outlines. These 2D sketches are combined with satellite imagery, publicly available photographs, archival records, and open datasets to generate a massing-level 3D model. Building height and volumetric characteristics are estimated using contextual cues such as surrounding structures, known architectural typologies, and scale references derived from people or urban elements. The resulting Digital Twin prioritizes geometric plausibility over fine architectural detail, enabling simulation, analysis, and decision-support tasks, such as environmental modeling, airflow and CFD approximation, and high-level Heritage BIM integration, while fully respecting cultural heritage restrictions. Three case studies illustrate the proposed workflow and systematically identify which components of conventional smart-building and Digital Twin pipelines remain feasible and which become infeasible under heritage regulations. The results demonstrate a practical and scalable path toward compliant Digital Twins for protected buildings, positioning low-detail modeling not as a limitation but as a regulatory necessity.

The rapid growth of the Internet of Things (IoT), including both consumer oriented and Industrial Internet of Things (IIoT) platforms, has given rise to a highly heterogeneous IoT platform landscape with significant differences in scalability, replicability, interoperability, deployment flexibility, and operation complexity, and hence selecting the most appropriate platform is a crucial issue in industrial and sustainability oriented engineering domains. This paper proposes a comprehensive and evidence based framework called ChriSmA2 to facilitate the systematic identification and evaluation of IoT platforms concerning their real world applicability, scalability, and replicability. The proposed framework is based on the application of PRISMA principles to a platform oriented evaluation process and can be used to perform a transparent and replicable screening and evaluation of both open source and proprietary IoT platforms. A comprehensive and unified set of evaluation criteria is defined for the assessment of IoT platforms from a multi dimensional perspective, such as community support, documentation quality, security features, deployment flexibility, platform maturity level, scalability, replicability, application domains, and operation limitations. The effectiveness of the proposed framework is validated through a large scale case study involving the systematic screening of a total of 255 candidate platforms and the subsequent evaluation of eighteen high quality platforms from different application domains such as smart cities, industrial automation, health care, and energy management. The evaluation results show a clear trade off between different types of platforms: while open source platforms show promising scalability and replicability due to the presence of a community driven and open nature of the platforms, they may face problems such as integration complexity and long term stability, whereas proprietary platforms may show better integration flexibility and operation robustness but may suffer from a lack of openness and flexibility. By providing a comprehensive and evidence based decision support framework through the synthesis of both technical and economic evidence, the proposed framework can guide engineers and decision makers to select sustainable and scalable IoT and IIoT platforms.

AR applications are rapidly gaining adoption, expanding users’ perceptual capabilities. Their potential for addressing visual occlusions and effectively extending the user’s line of sight significantly enhances their value in various contexts, from professional to personal use. This work presents a novel system designed to project 3D human poses onto AR glasses, enabling users to perceive concealed individuals behind solid objects, addressing a critical limitation of traditional visual perception. To achieve real-time and accurate 3D projection, we employ fiducial markers strategically placed within the environment. The markers are periodically fused with IMU sensor data to accurately estimate the user’s head orientation, a crucial step for correct spatial alignment. Furthermore, we leverage a multi-view 3D human pose estimation method using calibrated cameras and incorporate attention mechanisms. These mechanisms focus the system on relevant features, improving accuracy and minimizing 3D joint error. Our experiments demonstrate that the proposed framework accurately projects 3D skeletal representations onto AR glasses, even when significant occlusions are caused by solid objects or other occupants within the scene. This novel approach offers a method to enhance situational awareness in dynamic environments where visibility is compromised, potentially benefiting various applications, from first response scenarios to security and surveillance.

With the rapid growth of video content across domains such as live streaming, remote education, surveillance, and virtual meetings, there is an increasing need for intelligent systems that can automatically analyze and summarize key events from video data. This paper presents an end-to-end framework that leverages 5G connectivity, Automatic Speech Recognition (ASR), Optical Character Recognition (OCR), and Large Language Models (LLMs) to generate concise, context-aware event summaries in the form of video highlights and textual overviews, eliminating the need to watch full videos. The system calibrates energy thresholds for ambient noise, extracts speech, and identifies keyframes for OCR, using HSV-based color histogram differencing. Redundant OCR outputs are filtered using a Levenshtein-based similarity threshold to retain unique content. TextRank is applied to reduce token load before invoking the OpenAI API for abstractive summarization, with per-minute summaries providing high-level insights. These summaries are overlaid by an Augmented Reality (AR) application, powered by a Network Application that receives streaming, performs instance segmentation to identify speakers and employs a decision-making module that determines the placement of the summaries within the user’s field of view. Experimental evaluation on a private 5G testbed demonstrates smooth 4K streaming with low latency, stable Channel Quality Indicator (CQI), and no frame loss, confirming the system’s scalability and effectiveness in dynamic network conditions.