In recent years, regional bus operations have faced critical challenges including declining ridership and driver short-ages. To address these issues and develop efficient operational plans, collecting Origin-Destination (OD) data is essential. However, conventional OD data collection methods are constrained by labor costs and time limitations, while IC card-based approaches miss flat-rate users and cash-paying passengers. Additionally, the introduction of facial recognition cameras raises social acceptability concerns due to surveillance implications. To address these challenges, this study focuses on using Bluetooth Low Energy (BLE) advertising packets emitted by passengers' mobile devices. However, modern devices periodically and randomly change their BLE MAC addresses, making continuous tracking difficult throughout a trip. To overcome this, we propose a bus OD data estimation method that applies MAC address carry-over technique. This approach exploits the asynchronous nature of MAC address changes and the embedded information in packets to enable continuous tracking. To evaluate the effectiveness of the proposed method, we conducted field experiments on two different operational bus routes. The results demonstrate that the proposed method can estimate OD data with an average precision of $\text{7 0 \%}(\text{m a x}: \text{1 0 0 \%}, \text{m i n}: \text{5 3 \%})$ and an average recall of $\text{3 4 \%}($ max: $\text{5 9 \%}, \min: \text{1 5 \%})$.
This study proposes a system in which a tail-shaped device responds to smartphone message notifications by conveying emotions intuitively. Messages are analyzed for sentiment and classified into three polarities: positive, neutral, and negative. Corresponding tail movements are generated based on the results. A user study confirmed that participants could intuitively interpret the emotions from the tail's movement, demonstrating the effectiveness of this non-verbal notification method.
Many studies have investigated remote robotic arm operation systems using virtual reality environments. However, conventional VR teleoperation systems suffer from communication latency, which makes it difficult for operators to receive immediate visual feedback of the robot's posture. This mismatch between input and visual feedback may increase cognitive and psychological load and degrade task performance. We propose a teleoperation system that overlays a virtual robotic arm to provide immediate visual feedback of user operation prior to receiving the delayed video from the physical robotic arm. By superimposing a virtual robotic arm on the stereo video, the system enables operators to simultaneously maintain immersion in the remote environment and perceive the input posture of the robotic arm. We evaluated information presentation methods through remote robotic arm task experiments in VR and user feedback. The results confirmed that, under high-latency conditions, the virtual robotic arm display reduced operator anxiety and helped maintain confidence during teleoperation.
Reconfigurable Intelligent Surfaces (RIS) have recently emerged as a key enabling technology for future 6 G networks, offering the capability to manipulate the wireless propagation environment intelligently. Here, we investigate the use of multiple RIS panels to establish Line-of-Sight (LOS) connectivity between a transmitter and a receiver via successive signal reflections. A core challenge lies in selecting the optimal RIS-assisted reflection path that maximizes end-to-end channel quality. To address this, we formulate the path selection problem as a Mixed-Integer Programming (MIP) optimization task, where the cost of each link depends on the selection of preceding links, introducing significant computational complexity. Simulation results demonstrate that the proposed scheme achieves near-optimal performance with minimal loss, validating its practical feasibility. Our results indicate minimal performance loss using Follow-The-Leader (FTL) and Follow-the-Perturbed-Leader (FTPL) schemes.
This paper presents an adaptive user state estimation/detection system for indoor occupants using off-the-shelf Bluetooth (BT) and Bluetooth Low Energy (BLE) devices. The proposed system classifies the user states into occupancy, distance, and activity level, and estimates each by processing Received Signal Strength (RSS) from nearby devices. In this study, a two-stage channel adaptation approach is proposed: a generalized base model is applied initially, followed by automatic adaptation to an environment-specific model after accumulating real-time data. User presence is inferred using a signal continuity based method that tracks both connection status and interruption patterns. For distance estimation, we propose a Hybrid Nakagami Fading model that integrates path loss, shadowing, and multipath effects into a unified formulation. Activity recognition is performed using RSS based pattern modeling, further enhanced by clustering with Dynamic Time Warping (DTW) to improve robustness across user behavior patterns. Experiments conducted across seminar rooms, offices, and residential apartment demonstrate strong performance: 100 % presence detection accuracy with graceperiod handling, 0.41 m average error in distance estimation, and 81.25% activity recognition accuracy, showing that the proposed system can estimate the user state with high accuracy.
This study investigates how Ultra-Wideband (UWB) node orientation and smartphone posture affect indoor localization accuracy, revealing that certain configurations can significantly degrade performance in real-world settings. Classification models using only distance features detected problematic poses. These results reveal several challenges in smartphone-based UWB localization. Future work will focus on error modeling and compensation to improve its robustness and accuracy in everyday scenarios.
To mitigate load imbalance, which may lead to increased response time in private cloud environments, we present a response time prediction approach utilizing an eXtreme Gradient Boosting (XGBoost) model, alongside three dynamic load-balancing optimization models. These methods are designed to optimize both response time and resource allocation. Experimental evaluations conducted on a small-scale testbed demonstrate that the XGBoost model achieves the lowest response time prediction error compared to the benchmark methods. Among the load-balancing approaches considered in this study, including the proposed optimal models, our response time minimization model, designated as Optimization Model 1, outperforms the others in terms of delay and resource utilization. Although the Double-DQN model exhibits greater adaptability to environmental dynamics, it incurs significantly higher computation time.
While evacuation on foot is generally recommended during large-scale disasters, vehicular evacuation is permitted in imminent threats such as tsunamis. However, evacuation by car can cause traffic congestion. Providing drivers with appropriate evacuation routes based on traffic congestion is expected to be useful for rapid evacuation. As communication infrastructures may be unavailable during disasters, cellular networks may also be unavailable. Therefore, in this paper, we evaluate the effect of providing drivers with evacuation assistance information using vehicle-to-vehicle communication (V2V) and a Delay/Disruption Tolerant Network (DTN). The system's effectiveness depends not only on the positions of vehicles and shelters but also on drivers' psychological traits. We designed a driver's decisionmaking model considering their cognitive bias and evaluated the effect on the evacuation completion time. The simulation results illustrated that cognitive bias affects the effectiveness of the evacuation assistance method based on V 2 V and DTN.
In this paper, we propose a smartphone-based system that estimates the risk of indoor furniture tipping over in real time due to earthquakes. Large-scale earthquake disasters occur all over the world. Although attention tends to be focused on damage caused by tsunamis and building collapses, many casualties have also been caused by furniture collapses in the home. Therefore, we propose an app that recognizes dangers in the home and notifies users of dangerous spots. Specifically, furniture is identified by learning per-frame depth maps taken with a monocular RGB camera via the smartphone's ARCore Depth API and images of furniture classes. The geometric dimensions recovered from the depth are used to calculate a normalized risk score ranging from 0 (safe) to 1 (high risk) for each piece of furniture using a tipping probability model defined by the Architectural Institute of Japan. As a result, high-risk furniture can be highlighted on the bird's-eye view, and recommended anchoring measures are also shown.
PrioV2X is a software-defined, AI-enhanced emergency alert system that leverages event-driven platform Solace PubSub+ as a real-time resource for V2X communication. The system introduces dynamic topic routing, AI severity classification, and urgency-aware suppression through a dynamic Time-To-Live (TTL) mechanism. This approach enables low-latency dissemination of high-priority alerts while preventing message flooding and redundant alert propagation. Emergency alerts are given priority among thousands of emergencies using an ML model, through Solace PubSub+ locally deployed Docker Broker. The hierarchy topic structure supports scalable and filterable message flows in real-time. PrioV2X is evaluated using 20,000 real-world accident records and benchmarked against PrioMQTT and J2H-based systems, demonstrating superior performance in terms of reliability, throughput efficiency, and prioritization accuracy. The architecture presents a robust alternative to traditional cloud-reliant or REST-based alerting systems, prioritizing safety is of utmost importance.
As real-time security and adaptive network control become critical in intelligent moving environments, centralized security frameworks like Interface to Network Security Functions (I2NSF) face latency challenges. To address this, we propose the Interface to In-Network Computing Functions (I2ICF) framework, which enables intent-based policy management by relocating security functions and translation closer to the data path. I2ICF performs policy translation, enforcement, and monitoring at the edge or within mobile entities, reducing communication delays. We implement I2ICF on a Kubernetes-based testbed and demonstrate its feasibility in orchestrating service functions for intelligent moving objects (IMOs). These results suggest that I2ICF can serve as a foundation for low-latency, intent-driven management in dynamic environments.
During Reinforcement learning (RL) training, Exploration is the process that nudges the learning agent out of the comfort zone of its acquired environmental knowledge (exploitation) towards new states exploration. A right balance between exploitation and exploration accelerates convergence and helps avoid sub-optimal solutions. The classical methods $\epsilon$-greedy, uses an independent decremented value to strike a balance between exploration and exploitation. $\epsilon$-greedy achieves good results despite its blind control of the exploration- exploitation balance. In this paper, we propose a methodology of controlling the exploration-exploitation balance with the help of the reward value; the reward-contrast method. The reward value gives a feedback measure of the agent's action in the environment; hence it carries much information about the setting in which the learning is performed. Leveraging environment dynamics to infer the necessity of more exploration (or absence of it) avoids blind exploration strategies. The results of the method in various test environments provided improved results in terms of average rewards and time of convergence compared to the classical method.
This study proposes an innovative generalization method for reconstructing depth images of moving objects using Wi-Fi channel state information (CSI) in unseen environments. This method supports the widespread use of Wi-Fi imaging in scenarios such as security and elder care The existing Wi-Fi depth imaging model suffers from poor generalization to new environments due to the change of CSI which is variant of different indoor structures. To overcome this obstacle, we propose a training strategy composed of cross-modal supervision and metric learning for a deep-learning model based on teacher-student learning and variational autoencoder (VAE), to help the model adapt to new environments without requiring the ground truth of depth images. The cross-modal supervision utilizes a limited number of RGB images captured in the target environment for adaptation. The metric learning strategy searches for the most similar sample for the target data in the source domain minibatches based on the image features. The model is aided by auxiliary tasks of learning the core components of the depth, position, and shape of the target. Instead of using depth images, such components are calculated from the RGB images collected in the target environment via existing image processing methods, significantly improving the generalization of Wi-Fi depth imaging model and alleviating the burden of data collection.
The effectiveness of Internet of Things (IoT) applications often relies on access to diverse and high-quality time-series sensor data. However, real-world data is often limited by noise, imbalance, and insufficient coverage of rare or extreme conditions. This paper presents a preliminary study on conditional generative modeling for synthesizing realistic and controllable IoT time-series data. We outline a verification methodology comparing three representative models-TimeVAE, Time-Transformer AAE, and GuidedDiffTime-under a unified evaluation framework using discriminative, predictive, and visualization-based metrics. Our aim is to explore the potential of conditional generation in capturing contextual dependencies while maintaining temporal dynamics. This work serves as a foundation for future development of robust and contextaware data generation frameworks to support IoT applications such as anomaly detection and system testing.
This paper proposes two novel localization methods for Shared Augmented Reality (Shared AR) addressing limitations of conventional approaches. Traditional methods struggle when image features cannot be reliably detected or when users are located far apart, making spatial alignment difficult. To overcome these issues, a Geolocation-Based method and a Depth-Based method tailored for such challenging scenarios are introduced. Additionally, this paper proposes a method to improve geographic accuracy by leveraging a Shared AR network. The experimental results demonstrate that the proposed localization methods maintain practically acceptable error levels and are effective even in scenarios where conventional methods are inapplicable.
Recent advancements in wearable human activity recognition have increasingly leveraged multiple IMU (Inertial Measurement Unit) sensor devices. However, this has raised serious concerns regarding privacy risks, including the possibility of personal identification from raw IMU data and exposure during wireless transmission. In multi-device environments, the richness of sensor information enables the identification of individuals with high accuracy. To address the trade-off between privacy protection and utility in activity recognition, this paper proposes MultiAAE, a framework that combines an Anonymizing Autoencoder (AAE) with Adaptive Differential Pulse Code Modulation (ADPCM). Each sensor device performs lightweight, localized anonymization using an AAE tailored to its modality, and the output is further processed with ADPCM to enhance the overall anonymity of aggregated data while reducing communication cost. We implement the proposed method on an ESP32 microcontroller and evaluate it in terms of anonymization accuracy, power consumption, latency, and communication overhead. Experimental results show that MultiAAE outperforms conventional approaches in both privacy protection and resource efficiency, under both single-device and multi-device configurations.
Human Activity Recognition (HAR) using acoustic information is attracting attention because it is less intrusive and better preserves privacy. However, developing generalized activity recognition models that perform reliably across diverse households remains challenging. In this study, we propose an acoustic-based daily living activity recognition system that leverages CLAP-based audio embeddings combined with a lightweight supervised learning model. We collected a new acoustic dataset comprising 23 daily in-home activities (e.g., drinking, working on a PC, watching TV), recorded from 16 participants. We extracted acoustic features using audio embeddings from a CLAP model pretrained via text-audio contrastive learning, and trained a 1D-CNN classifier using the extracted features. We evaluated the model using four validation strategies: hold-out, stratified hold-out, leave-one-group-out (LOGO), and leave-one-participant-out (LOPO). Furthermore, we conducted zero-shot classification using CLAP, which achieved an accuracy of 30%, highlighting the challenge of recognizing complex daily activities without task-specific training. In contrast, supervised learning with CLAP-based features achieved a LOPO accuracy of $77.99 \% \pm 8.06 \%$, demonstrating strong generalization across unseen participants. Additionally, t-SNE analysis revealed that CLAP embeddings formed more compact and distinct clusters than traditional spectrogram features, demonstrating their superior discriminative power.
Intent-based systems empower autonomous networks to operate based on high-level declarative goals, offering flexibility, scalability, and reduced operational complexity. While the security implications of such systems are recognized, intent injection attacks, which adversaries manipulate syntactically valid but semantically malicious intents, have remained largely theoretical due to the lack of standardized intent samples for empirical evaluation. This paper presents a semantic-level injection detection framework that employs a fine-tuned BERT classifier to transform each intent into a natural language representation. This enables semantic analysis of the intent's content and classifies it as either benign or injected. To bridge the gap of unavailable real-world data, we synthesize training samples guided by benign and adversarial intent policies. For evaluation, we use intent samples from the recently released ETSI 3GPP TS 28.312 specification, creating variants by injecting code into their YAML structures to preserve structural validity while altering semantic meaning. Experimental results show that the proposed method effectively detects injected intents, achieving an AUC of 0.98 and an F1-score of 0.97, significantly outperforming existing injection detection methods such as PromptGuard and VulBERT, which prove ineffective against these attacks. These results underscore the necessity of developing tailored detection mechanisms for intent-based systems and highlight the potential of semantics-aware approaches to address this unknown known security threats.
This paper presents a Hybrid Indoor Localization Scheme (HILS) for drones operating in indoor environments, utilizing Ultra-Wideband (UWB) technology to provision a high positioning accuracy to the drones. The proposed system combines Time Difference of Arrival (TDoA) for initial absolute position estimation and Two-Way Ranging (TWR) for precise relative distance measurements between a pair of drones. After computing their absolute positions via TDoA, the inter-drone distances derived from these positions are compared against TWR-measured distances. The discrepancies are iteratively reduced by particle filtering to enhance the positional accuracy of drones in indoor environments. This fusion approach leverages the complementary strengths of TDoA and TWR to reduce localization errors. Experimental results in simulation demonstrate that HILS significantly improves the localization accuracy of a scalable TDoA-based system.
Traffic congestion caused by an excessive influx of tourists into sightseeing areas has become a serious issue. In this paper, the authors propose a heuristic route guidance method that individually provides incentives to tourists to encourage detours and reduce traffic congestion. From the results of multiagent simulation using artisoc Cloud, the authors confirmed that the proposed method can guide tourists to less congested routes with less total incentives and computation time compared to the route guidance method using random search (RS) and genetic algorithm (GA).
