Special issue on smart city technologies and services based on AI for digital twin applications

Abstract

The rapid advancement of artificial intelligence (AI) technologies, along with the accelerated development of smart cities, has created unprecedented opportunities to enhance the efficiency and sustainability of urban environments. The 2022 special issue on smart cities focused on foundational machine learning (ML), Internet of Things (IoT)-driven analytics, and optimization techniques to improve traffic management, public safety, and urban infrastructure sharing [1]. Smart cities go beyond basic digitalization by incorporating data-driven decision-making and advanced automation to improve citizens' quality of life, reduce energy consumption, and address urban challenges such as traffic congestion and environmental degradation. Since then, emerging trends such as metaverse integration, privacy-preserving AI, edge AI, and LiDAR-based autonomous navigation have reshaped smart city applications. At the same time, these advancements pose complex challenges that require integrated technological and governance strategies.

Recent research on smart city development has focused on integrating high-level intelligence into urban systems and analyzing the economic ripple effects of these technologies on wider industrial ecosystems. Massive datasets generated in real time through sensors and IoT devices provide critical insights into traffic flows, environmental conditions, energy usage, and patterns of human activity. However, the transformation of these large-scale datasets into actionable intelligence remains a significant technical and managerial challenge.

To address these challenges, the convergence of AI and digital twin technologies has emerged as a promising solution. This convergence enables the integration and analysis of heterogeneous data sources, offering predictive insights and real-time decision-making capabilities that enhance operational efficiency, optimize resource utilization, and strengthen the sustainability and resilience of urban systems. The applications of AI in smart cities span a wide range of domains, including anomaly detection, traffic flow analysis, predictive maintenance, energy optimization, and public safety. When combined with robust data privacy and security frameworks, AI can support transparent and accountable governance and safeguard personal information.

Digital twins are dynamic virtual models of physical urban environments that enable simulation-based policy testing and proactive problem resolution. These models allow city administrators to simulate infrastructure scenarios, forecast outcomes, and manage assets. When augmented with AI, digital twins achieve more precise feature extraction, automated fault detection, and scalable predictive analysis, which in turn yield cost savings and operational improvements. Furthermore, the fusion of digital twin technologies with metaverse platforms creates immersive and interactive environments to enable citizens to engage and contribute to urban planning. Such integration not only promotes participatory governance and democratizes decision-making but also enhances citizen trust and engagement in smart city initiatives.

Against this backdrop, the Electronics and Telecommunications Research Institute (ETRI) Journal has organized this special issue to present state-of-the-art research and practical applications that explore the synergy between AI and digital twin technologies for the advancement of smart and sustainable urban ecosystems. Contributions were solicited from academia, research institutions, and industry professionals, and all submissions underwent a rigorous peer-review process. As a result, seven high-quality papers were selected for inclusion in this issue, covering a broad range of topics including sewer infrastructure management, LiDAR odometry, urban traffic datasets, occupancy sensing, GPU sharing strategies, fault detection methods, and virtual try-on systems.

The following sections introduce the key contributions of each selected paper and highlight their significance in shaping the future of intelligent and sustainable urban development.

The first paper [2], titled “Trends in Intelligent Sensor-Based Customized Management Technologies for Sewer Infrastructures,” by Kang and others, offers a comprehensive overview of intelligent sensor-based sewer management technologies, identifying both opportunities and challenges and contributing to the advancement of sustainable and efficient sewer infrastructure systems. This paper examines the potential applications and associated challenges, including IoT-driven data collection, machine learning and deep learning analytics, cloud and edge computing, and autonomous robotics. Based on case studies from South Korea, Germany, Japan, France, Singapore, the United Kingdom, and the United States, this paper highlights the effectiveness of digital twins, real-time monitoring, and predictive maintenance, as well as persistent challenges such as sensor durability, robotic mobility, and data analysis limitations. By providing a foundation for technological innovation, this study proposes strategies and roadmaps to ensure the stable adoption and continuous development of smart sewer management systems.

In the second paper [3], titled “ELiOT: END-to-End LiDAR Odometry with Transformers Harnessing Real-World, Simulated, and Digital Twin” by Lee and others, propose ELiOT, a transformer-based LiDAR odometry framework that integrates real-world, simulated, and digital twin data for training. This study introduces a method that leverages a 3D transformer and a self-attention-based flow-embedding network to enable accurate urban navigation while effectively bridging the domain gap between simulation and real-world environments.

The third paper [4], titled “DOROS: A Multi-Level Traffic Dataset for Dynamic Urban Scene Understanding” by Kang and others, addresses the pressing need for diverse and richly annotated datasets in smart traffic systems. Whereas existing datasets often provide only limited scene annotations and lack sufficient diversity across traffic conditions, weather, and locations, the authors present DOROS, a large-scale dataset comprising 49,296 images. It provides structured annotations across agent, location, and behavioral categories, offering a comprehensive resource for understanding complex urban scenes. To demonstrate its difficulty and utility, the authors benchmarked the dataset using both widely adopted convolutional neural network (CNN)- and Transformer-based object-detection models. This dataset is expected to be a valuable resource for researchers working on autonomous driving, traffic management, and digital twin applications in smart cities.

The fourth paper [5], titled “Privacy-Preserving Labeling-Free Occupancy Counting Sensor Based on ToF Camera and Clustering” by Jeong and others, addresses the challenge of occupancy detection in smart buildings, where conventional camera-based approaches often raise privacy concerns. To overcome this issue, the authors leveraged Time-of-Flight (ToF) cameras instead of red, green, and blue (RGB) imaging and applied a traditional clustering technique to detect occupants without the need for labeled data. Experimental results demonstrate that the proposed method achieves over 90% accuracy in single-entry scenarios and delivers superior performance compared to deep-learning-based object detection methods. This study is expected to contribute significantly to privacy-conscious building monitoring and digital twin-driven energy management.

The fifth paper [6], titled “ Exploring GPU Sharing Techniques for Edge AI Smart City Applications” by Woo and others, investigates GPU sharing strategies to support efficient edge AI in smart city applications such as traffic management, surveillance, and environmental monitoring. Using the NVIDIA Jetson AGX Orin platform and YOLOv8 workloads, the study compares threading and multiprocessing approaches, showing clear tradeoffs between memory usage and inference speed. While threading reduces memory consumption by sharing CUDA contexts, multiprocessing achieves higher GPU utilization and faster inference. The paper also highlights scalability issues related to synchronization overhead and resource contention.

In the sixth paper [7], titled “Robust Mahalanobis Distance-Based Lazy Learning Method for Fault Detection in High-Dimensional Processes” by Yu and others, the authors address the challenge of fault detection in high-dimensional processes, where traditional Mahalanobis distance (MD)-based methods suffer from increased type I errors owing to the curse of dimensionality. This study highlights how sparse data regions in high-dimensional spaces cause unstable covariance matrix estimations, undermining the reliability of classical MD approaches. To overcome this problem, the authors propose a robust MD-based lazy-learning method that employs the minimum covariance determinant technique to estimate robust covariance matrices. This method is integrated with baseline learners, such as k-nearest neighbors and local outlier factors, but is broadly applicable to other lazy-learning approaches. Experimental validation of the benchmark processes demonstrates that the proposed method significantly improves the fault detection performance, effectively reducing type I errors in high-dimensional settings.

The seventh paper [8], titled “High-Speed and Precise Virtual Try-On with Two-Stage Semantic Segmentation and Latent Consistency Model for Optimized Diffusion Processes” by Baek and others, examines whether the segmentation mask accuracy, rather than the generative model, is the key limitation in current virtual try-on (VTON) systems. The authors propose HSP-VTON, a framework that combines a refined two-stage semantic segmentation approach to improve the mask precision with a Latent Consistency Model that accelerates diffusion-based image generation. This integration directly addresses the dual challenges of achieving high-quality garment alignment and reducing the computational cost. Experiments on the ATR dataset demonstrate a 2.8% improvement in the mean Intersection over Union (mIoU), whereas evaluations on VITON-HD demonstrate superior LPIPS and SSIM performance over state-of-the-art models. Additionally, the proposed approach reduces the number of diffusion inference steps from 30 to 5, substantially reducing the processing time without compromising visual quality.

The guest editors would like to thank all authors, reviewers, and editorial staff of the ETRI Journal for making this special issue successful. We are pleased to have contributed to the effort to present high-quality technical papers in a timely manner. These studies represent leading research on smart city technologies and services, particularly emphasizing the role of AI in enabling digital twin applications.

The authors declare that there are no conflicts of interest.