Junho Jeong

Fall detection in multi-person environments, such as nursing homes and rehabilitation centers, is essential for ensuring the safety of vulnerable populations. Despite advances in deep learning, current vision- and skeleton-based fall detection systems often exhibit false negatives and reduced reliability in real-world scenarios due to scene complexity. This study presents a system-level analysis of fall detection errors by comparing four approaches—two skeleton-based methods using ST-GCN and ProtoGCN, a rule-based method, and a VIRA-GCN-based 3D joint method—on 95 RGB video sequences captured under minimally constrained multi-person conditions. We define six error types: skeleton structural interference, localized joint recognition failure, temporal skeleton identity inconsistency, object-to-skeleton association failure, viewpoint-induced observation limitation, and action-level ambiguity with similar activities. Although most methods achieved high event-level recall, their false-positive and false-negative patterns differed. The rule-based approach showed the most balanced performance, under the present experimental conditions, whereas the ST-GCN-based skeleton approach was more sensitive to joint-level and tracking instability. ProtoGCN reduced false positives but increased false negatives, showing a more conservative decision pattern. The VIRA-GCN-based 3D joint approach provided spatial cues but did not eliminate upstream pose and tracking errors. These results highlight the need for skeleton–depth fusion, robust identity tracking, occlusion handling, and enhanced joint recognition in real-world multi-person fall detection.

Ship detection is essential for safe maritime and inland-waterway navigation and port logistics, yet adverse weather (rain, snow, fog) degrades visibility and increases missed detections and false alarms. This study proposes Robust Ship-detection YOLO for Adverse Weather (RSYA), an enhanced real-time detector for maritime surveillance in degraded environments. As an artificial intelligence contribution, RSYA augments a You Only Look Once version 11 nano (YOLO11n) baseline by adding a high-resolution P2 detection head for small vessels, introducing C3k2-Triplet blocks to strengthen feature representation, and using Focal-Complete Intersection over Union loss to emphasize hard, small targets. As an engineering application, we generate and evaluate on adverse-weather ship imagery with controlled rain, snow, and fog corruption, together with corresponding original-weather data. Across comparative experiments, RSYA improves robustness while preserving real-time throughput. On the adverse-weather dataset, RSYA increases recall by 2.90%, mean Average Precision at Intersection over Union 0.5 (mAP_{50}) by 2.29%, and mean Average Precision averaged over thresholds 0.5–0.95 (mAP_{50-95}) by 3.35% over the YOLO11n baseline. These results show that combining weather-oriented data synthesis with targeted architectural and loss redesign substantially improves ship-detection performance under degraded visibility.

Falls remain a critical safety concern in surveillance settings, yet monocular RGB methods often degrade in multi-person scenes with occlusion and loss of three-dimensional cues. This study proposes a neuro-symbolic framework that restores physically interpretable depth proxies from monocular video and fuses them with skeleton-based spatio-temporal inference for robust fall detection. The pipeline estimates per-frame depth and 2D skeletons, recovers world coordinates for key joints, and derives absolute neck height and vertical descent rate for rule-based adjudication, while a neural method operates on joint trajectories; final decisions combine both streams with a logical policy and short-horizon temporal consistency. Experiments in a realistic indoor testbed with multi-person activity compare three configurations—neural, symbolic, and fused. The fused neuro-symbolic method achieved an accuracy of 0.88 and an F1 score of 0.76 on the real surveillance test set, outperforming the neural method alone (accuracy 0.81, F1 0.64) and the symbolic method alone (accuracy 0.77, F1 0.35). Gains arise from complementary error profiles: depth-derived, rule-based cues suppress spurious positives on non-fall frames, while the neural stream recovers true falls near rule boundaries. These findings indicate that integrating monocular depth proxies with interpretable rules improves reliability without additional sensors, supporting deployment in complex, multi-person surveillance environments.

Recently, research utilizing smart home technology for anomaly detection in single-person households has been actively conducted. However, real-world sensor data flaws, such as noise and battery discharge, cause the 'Garbage In, Garbage Out (GIGO)' problem, thereby undermining model reliability. Furthermore, existing research on fixed sensors faces limitations in logically explaining complex life patterns, such as 'sleep', as it relies solely on simple activation status. To address these issues and lay the foundation for anomaly detection based on daily life patterns, this study proposes a robust extraction methodology that fuses heterogeneous sensors including respiration, heart rate and illuminance and transforms raw data into 'Interpretable Representations', referred to as daily life patterns such as sleep, outing, and activity. Specifically, we construct these patterns by determining dynamic thresholds via K-Means clustering and comprehensively considering global configuration of all sensors. Case Study results demonstrate that the proposed method resolves the GIGO problem by clearly distinguishing 'long-term absences' from 'long sleep' and accurately differentiating actual missing data marked as ' NaN ' from non-occurrences represented as ' 0 ', thereby enabling the extraction of robust daily life patterns.

Despite advancements in deep learning-based Monocular Depth Estimation (MDE), applying these models to video sequences remains challenging due to geometric ambiguities in texture-less regions and temporal instability caused by independent per-frame inference. To address these limitations, we propose STF-Depth, a novel post-processing framework that enhances depth quality by logically fusing heterogeneous information—geometric, semantic, and panoptic—without requiring additional retraining. Our approach introduces a robust RANSAC-based Vanishing Point Estimation to guide Dynamic Depth Gradient Correction for background separation, alongside Adaptive Instance Re-ordering to clarify occlusion relationships. Experimental results on the KITTI, NYU Depth V2, and TartanAir datasets demonstrate that STF-Depth functions as a universal plug-and-play module. Notably, it achieved a 25.7% reduction in Absolute Relative error (AbsRel) and significantly enhanced temporal consistency compared to state-of-the-art backbone models. These findings confirm the framework’s practicality for real-world applications requiring geometric precision and video stability, such as autonomous driving, robotics, and augmented reality (AR).

Graph Neural Networks (GNNs) have recently achieved remarkable success in various learning tasks involving graph-structured data. However, their application to multi-relational graph anomaly detection problems on real-world datasets presents several challenges that significantly hinder performance: the structural noise and inconsistencies inherent in real-world graph data, difficulties in aggregating information across multiple relation types, and the imbalanced class labels. To address these limitations, we introduce a graph structure learning layer designed to refine the original, noisy graph structure, enhancing the representation of node relationships. This enables our model to effectively handle inconsistencies and structural noise present in the original graph data with the augmented graph topology. To further exploit the semantic information embedded within the multi-relational graph data, we developed a multi-scale aggregation layer that facilitates information propagation and fusion through multi-relational levels, utilizing automatically and implicitly extracted meta-path graphs. Additionally, to handle the impact of imbalanced class labels, which often degrades the classification performance of GNN-based approaches, we introduce a score reweighting mechanism that balances probability scores between minority and majority classes. This strategy improves the model’s learning ability, leading to improved class distinction. Comprehensive experiments are conducted using two popular real-world review datasets to validate the effectiveness of our proposed model. Empirical results demonstrate that our approach consistently outperforms state-of-the-art baselines across three evaluation metrics, namely Average Precision, AUC and GMean.

In aerial surveillance using drones, person re-identification (ReID) is crucial for public safety. However, low resolutions in drone footage often leads to a significant drop in ReID performance of subjects. To investigate this issue, rather than relying solely on real-world datasets, we employed a synthetic dataset that systematically captures variations in drone altitude and distance. We also utilized an eXplainable Artificial Intelligence (XAI) framework to analyze how low resolutions affect ReID. Based on our findings, we propose a method that improves ReID accuracy by filtering out attributes that are not robust in low-resolution environments and retaining only those features that remain reliable. Experiments on the Market1501 dataset show a 6.59% percentage point improvement in accuracy at a 16% resolution scale. We further discuss the effectiveness of our approach in drone-based aerial surveillance systems under Fog/Edge Computing paradigms.

Agriculture is essential for human survival facing critical issues such as limited access to modern machinery, a lack of comprehensive data, and difficulties in implementing insurance programs. To address these challenges, this study explores the potential of blockchain technology, finding that the Hyperledger Fabric can satisfy certain criteria vital for rectifying the existing issues due to its secure and scalable features. Utilizing the channel mechanism of Hyperledger Fabric, we developed a multi-channel network infrastructure to cater to the varying needs in agriculture. Each agricultural service is facilitated through its unique channel, accompanied by an easy-to-navigate webpage, enhancing user experience while ensuring data security. In our experiments, we deployed the system using Docker containers and virtual machines. Performance tests were conducted to evaluate the system’s effectiveness in data sharing, machinery service coordination, and insurance claim processing. The results demonstrated that the system effectively improves data sharing efficiency, streamlines machinery service coordination, and enhances the transparency and reliability of insurance claims processing. These findings highlight the potential of our system to significantly enhance agricultural production efficiency. The flexible and expandable nature of this system offers tremendous potential for future improvements, aiming to fulfill the multifaceted demands of agriculture more effectively, thereby laying a robust foundation for significantly enhancing agricultural production efficiency.

Over the past decade, there has been a significant surge in consumer application services and server connectivity, and this trend is expected to double in 2030. The primary contributors to the increased demand for network resources are devices connected through third-party service providers and mobile operators. Many prominent consumer services rely on a client-server architecture, which can introduce latency delays in the communication channel. Additionally, peer-to-peer (P2P) communication places a substantial load on eHealth servers, leading to service delays. In this research paper, we propose a model for scheduling heterogeneous workloads and recommending resources for eHealth edgecloud connectivity using Federated Split Learning (FSL) model for smart cities. Distributed FSL offers a robust solution for handling both direct and indirect user requests through a distributed mobile core operator stack. This technique empowers eHealth administrators to locally learn optimal policies and make informed decisions by prioritizing resource allocation and scheduling. We demonstrate the effectiveness of this technique through an active simulation server designed for track-driven caching policy and local policy scheduling, ultimately enhancing resource recommendation in eHealth applications. The proposed technique is focused on the development of a heterogeneous workload recommendation system and obtained accuracy of 89.63% over 200 users trails.

In recent years, the convergence of edge computing and sensor technologies has become a pivotal frontier revolutionizing real-time data processing. In particular, the practice of data acquisition—which encompasses the collection of sensory information in the form of images and videos, followed by their transmission to a remote cloud infrastructure for subsequent analysis—has witnessed a notable surge in adoption. However, to ensure seamless real-time processing irrespective of the data volume being conveyed or the frequency of incoming requests, it is vital to proactively locate resources within the cloud infrastructure specifically tailored to data-processing tasks. Many studies have focused on the proactive prediction of resource demands through the use of deep learning algorithms, generating considerable interest in real-time data processing. Nonetheless, an inherent risk arises when relying solely on predictive resource allocation, as it can heighten the susceptibility to system failure. In this study, a framework that includes algorithms that periodically monitor resource requirements and dynamically adjust resource provisioning to match the actual demand is proposed. Under experimental conditions with the Bitbrains dataset, setting the network throughput to 300 kB/s and with a threshold of 80%, the proposed system provides a 99% performance improvement in terms of the autoscaling algorithm and requires only 0.43 ms of additional computational overhead compared to relying on a simple prediction model alone.