Hearables, wearable earphone-type devices, are often operated via voice commands. However, voice control faces limitations such as sensitivity to environmental noise, difficulty in silent environments, and privacy concerns. Silent speech interaction (SSI) has been proposed to address these issues, with methods using IMUs and ultrasound showing promise. Yet, SSI using ear canal pressure remains underexplored. In this study, we propose a novel SSI method that recognizes silent speech by analyzing pressure changes in the ear canal, measured by a built-in barometric sensor. These signals are classified using machine learning techniques such as SVM, DTW, and kNN. Experiments were conducted with seven participants using five predefined commands related to music control. The proposed method achieved an average recognition accuracy of 72.0
Future communication systems are complex infrastructures with virtualized software services that are updated regularly as requirements evolve. This paper addresses the adaptation of microservices under dependency and efficiency constraints. We formalize the evolution of microservices as a combination of two subproblems, the Microservice Dependency Problem and the Evolution Planning Problem, both of which are difficult to solve optimally. We then propose a method based on the Binate Covering Problem (BCP) with branch-and-bound, and introduce a novel algorithm that finds the deployment steps towards a desired new configuration. Our proposed method, DEP-DS, is then compared with two heuristics on three sample datasets from historical radio access network update records run on Kubernetes. We further show that BCP with greedy search is faster but finds fewer solutions to the evolution plan. Overall, DEP-DS is able to find solutions to all samples, generates deployment plans within an average time of 1–2 s, and the plans are similar to other heuristics in terms of CPU usage.
In this paper, we propose a song recommendation method for DJs that focuses on song lyrics. A DJ plays and connects a variety of songs. There are two ways to connect songs in DJ: based on musical similarities and based on semantic similarities. DJs of J-POP and anime songs tend to use connecting methods based on semantic similarity. In this paper, we focus on the lyrics of songs to support DJs of J-POP and anime songs. The task is to recommend the next song for the currently selected song. To train our model, we create a dataset from actual set lists played at DJ events. We fine-tune the Japanese pre-trained BERT in two ways: a Bi-Encoding model and a Cross-Encoding model. We experiment with Japanese pre-trained BERT as a baseline. In MRR, the Cross-Encoding model showed superior scores. In user tests, the Bi-Encoding model showed superior scores.
This study proposes a system to support the establishment of midfoot strike (MFS), a foot strike pattern where the heel and forefoot contact the ground simultaneously. While MFS is considered to reduce injury risk and improve running performance, it is difficult to acquire and maintain consistently. To promote MFS acquisition, runners must become aware of their ground contact balance and the distribution of plantar pressure during running. The proposed system enhances this awareness by providing intermittent visual feedback, which consists of insoles with pressure sensors placed at the heel and the ball of the foot and an LED mounted on the brim of a cap. When pressure values exceed a threshold, the LED lights up in red (heel) or blue (ball of the foot). Rather than delivering constant feedback, the system provides feedback at fixed intervals, encouraging users to estimate their foot strike internally and later verify its accuracy. This cycle of estimation and correction helps runners gradually learn to identify and correct deviations from MFS on their own. Even after the system is removed, users are expected to have learned to maintain MFS independently. An experiment was conducted to evaluate the system’s effectiveness. The results showed that participants who used the system could detect and correct deviations from MFS earlier than those who did not. This suggests that the intermittent feedback is effective for supporting the acquisition and retention of MFS during running.
Human skeleton data plays a crucial role in healthcare, rehabilitation, and human-computer interaction. However, traditional motion capture methods are often expensive and highly dependent on controlled environments. To address these limitations, we propose a novel framework based on Diffusion and Transformer models that reconstructs full-body skeleton sequences from foot pressure data collected via smart insoles.
Retrieving relevant information from long videos remains a significant challenge due to high computational costs, semantic redundancy, and the need for temporal reasoning. We propose LightVideoRAG, a lightweight retrieval framework tailored for long-video question answering. LightVideoRAG combines adaptive frame sampling, which filters out redundant frames while preserving key semantic content, with context-aware retrieval modules that integrate both local neighborhood signals and global temporal information. This design enables efficient temporal grounding without processing entire video sequences. Unlike existing methods that rely on dense captioning or proprietary APIs, our system operates entirely on a locally deployed Vision-Language Model (VLM), ensuring strong data privacy and low latency. Evaluations on the LongVideoBench and Video-MME benchmarks show that LightVideoRAG achieves substantial gains in QA accuracy while requiring only a fraction of the computational resources, outperforming the base model and approaching the performance of larger size models. This demonstrates its potential as a scalable and accessible solution for efficient video understanding in resource-constrained environments. Our code is available at https://github.com/linshys/lightvideoRAG .
Recently, with the improvement of computing power in mobile devices and the development of augmented reality (AR) technology, there has been growing interest in using image recognition for information retrieval and navigation. By using AR technology, systems that merge the real world and virtual space based on camera images and sensor data can provide users with intuitive information. However, many conventional systems require the installation of AR markers or beacons for their implementation and operation. It is important to design a highly versatile system with that does not require additional hardware or the preparation of special environments. In this paper, we propose an AR overlay display system with video synchronization. The proposed system compares camera images from Android devices with images registered in advance on a server created with FastAPI in real-time. In addition, this system can be easily operated within a server environment using devices in the same network, without the need for dedicated equipment or special environments. An evaluation of the proposed method showed that as the number of registered images increased, the amount of matching processing also increased and the execution time became longer. Furthermore, the system’s recognition rate was 100 0^∘ and 30^∘ , and images could be recognized by the system’s terminal at an average occlusion rate of 65
The use of Artificial Intelligence (AI) in the insurance industry covers a wide spectrum, including risk analysis, fraud detection, personalized policies, and customer support through chatbots. However, there is still room for improvement, as many techniques still rely heavily on manual processes. In this paper, we discuss both current solutions and a new system aimed at improving decision-making in the insurance field. We introduce a modular decision support system that weaves a Large Language Model (LLM) into crucial stages of the insurance decision-making process. Unlike fully automated systems, our approach embraces a Human-in-the-Loop (HITL) model, prioritizing transparency, user control, and explainability. This prototype allows insurance professionals to interact with data and AI tools using natural language, making it easier to handle tasks like portfolio analysis, customer segmentation, drafting personalized proposals, and querying structured data, without need for technical know-how.
In saxophone performance, the shape of the mouth and the use of breath are crucial factors. However, because these elements are not visually observable, mastering the instrument can be challenging. This study proposes a method for recognizing mouth movements during saxophone playing by focusing on changes in ear canal pressure. Since ear canal pressure varies depending on the positional relationship between the mandible and the ear canal, it serves as an effective non-contact means of measuring mouth movement. We developed a classification model to distinguish four representative types of mouth movements used in saxophone performance. Features were extracted from the pressure data and classified using base models including SVM, KNN, and Random Forest. Logistic regression was applied to integrate the outputs of these classifiers. An evaluation experiment involving five participants achieved classification F-scores ranging from 89.9
When calculating features in activity recognition, applying a fixed-length sliding window to time series data without considering the timing of activity transitions (hereinafter referred to as activity change points) may result in multiple activities being mixed within a single window, thereby reducing recognition accuracy. To address this issue, we propose a segmentation method that utilizes plantar pressure sensors placed on the foot to automatically detect moments of foot-ground contact. Since these contact moments often correspond to the boundaries between distinct lower-body movements, they can be used to segment continuous activity data into more homogeneous segments. This approach helps reduce the likelihood of mixed activities within a single analysis window. In our evaluation, we compared the proposed method with a conventional segmentation approach based on the spectral transition measure of acceleration data. Although the proposed method did not outperform the conventional method in terms of overall segmentation or activity recognition accuracy, it showed better segmentation performance in specific transitions characterized by distinct foot-ground contact patterns, such as transitions from sitting to walking or from walking to ascending stairs. These findings suggest that plantar pressure–based segmentation can serve as a valuable supplement to existing approaches, particularly in scenarios involving lower-limb activity transitions.
Napping is known to have many benefits for cognitive and psychological health, specifically improving alertness and mood and reducing stress. Relaxation prior to napping can enhance its effectiveness, and while auditory relaxation is commonly used, Virtual Reality (VR) offers an immersive environment that integrates visual and auditory senses, potentially producing greater relaxation effects. However, studies measuring the effects of VR-based audiovisual relaxation on napping remain limited. In our experiment, 12 participants experienced a nap with VR-based audiovisual relaxation and a nap with auditory relaxation. The results showed that napping with VR-based audiovisual relaxation significantly increased HRV scores, an objective indicator of stress, compared to napping with auditory relaxation, while no significant differences were found in alertness and mood. These findings suggest that VR-based relaxation may further enhance stress reduction during naps, warranting further research into its applications across various settings.
Mixed Reality (MR), which overlays virtual user interfaces (UIs) onto physical environments, necessitates UI adaptation to complex scenes and tasks. While prior work focused on adaptive spatial UI layout in MR, functional adaptation such as widget recommendation remains underexplored. We present a user study ( n = 16 ) using a Large Language Models (LLMs)-powered widget recommender system (RS) as a technology probe to investigate how context-aware recommendations affect user experience. The system uses LLMs with contextual data (reading text, video transcript and typed data) to suggest MR widgets. Results show that widget recommendations facilitated access to context-relevant functionalities and simplified task workflows, thus enhancing the user experience and reducing workload. However, the usability of the widget RS depends on appropriate widget design and recommendation strategies that enable personal customization. This study serves as an initial step toward MR widget RSs and offers insights for adaptive user-RS interactions.
When a sound source that can interfere with concentration exists in the same direction as a user’s focused task, it can lead to problems such as reduced ease of work and increased workload. This study addresses situations where a distracting sound source is present in the same direction as the user’s focused task and proposes a method to improve the ease of performing the task. The proposed method uses AR technology to virtually move only the visual information of the distracting sound source (e.g., the appearance of people conversing) to a direction different from the user’s focused task. This aims to make the user perceive the distracting sound source as being at that virtual location, thereby improving the ease of task performance. The results of an evaluation experiment confirmed that the proposed method has the effect of significantly reducing the workload of a cognitive task. These findings aid the design of AR-based task concentration support systems.
This paper presents human-centered design recommendations for agricultural software grounded in the practices of smallholder farmers in Central Borneo. Using a mixed-methods approach, which are survey (n = 286), interviews (n = 20), and co-design sessions (n = 20). From this study, we identified patterns of device use, connectivity, and information behavior, alongside themes of digital access, trust, and peer validation. Co-design activities further contextualized these findings, producing functional and non-functional requirements for agricultural applications. The resulting recommendations emphasize mobile-first, low-complexity interfaces, offline and data-efficient features, peer-driven knowledge exchange, localized content, and mechanisms for building trust. While exploratory and limited to one region, the study demonstrates how participatory methods can address socio-technical constraints and generate actionable requirements. It contributes to human-centered computing by showing how inclusive design processes inform the development of trustworthy and context-aware agricultural platforms for low-resource environments.
Protecting personal computers (PCs) from unauthorized access typically relies on password authentication, which is known to suffer from cognitive burden and weak credentials. As many users nowadays carry mobile devices with advanced security features throughout their day, there is an opportunity to leverage these devices to improve authentication to PCs. In this paper we empirically evaluate a token-based, passwordless approach in which users authenticate to their PC by confirming the authentication request on their smartphone or smartwatch. Upon a request to login to the PC, or to evaluate privileges, the PC issues an authentication request that users receive on their mobile devices, where users can confirm or deny the request. We evaluate button tap and biometric fingerprint verification as confirmation variants, and compare their authentication duration, success rate, and usability to traditional password-based authentication in a user study with 30 participants and a total of 1,200 authentication attempts. Smartwatch-based authentication outperformed password-based authentication and smartphone-based variants in authentication duration, while showing comparable success rates. Participants rated smartwatch-based authentication highest in usability, followed by password-based authentication and smartphone-based authentication.
Image sensors are integral components of many critical intelligent systems. However, a growing threat, known as Electromagnetic Signal Injection Attacks (ESIA), poses a significant risk to these systems. ESIA enables attackers to remotely manipulate images captured by cameras which can potentially lead to malicious actions and catastrophic consequences. Despite the severity of this threat, the effects of ESIA remain poorly understood, and effective countermeasures are lacking. This paper aims to address these gaps by investigating ESIA from two distinct aspects: pixel loss and color strips. By analyzing these aspects separately on image classification tasks, we gain a deeper understanding of how ESIA can compromise intelligent systems. Additionally, we explore a lightweight solution to mitigate the effects of ESIA. Our findings provide valuable insights for future research and development in the field of camera security and intelligent systems.
This study proposes estimating forearm pressure using a PPG sensor in a smartwatch or activity meter. The method involves creating a regression model that estimates the applied pressure by analyzing changes in the local maximum values of pulse waves before and after the pressure is applied. Experiments with five subjects evaluated the method. The individual evaluation model, using only the user's data, had an average coefficient of determination of 0.57 and an MAE of 1.60 Kg. The total evaluation model, using all subjects' data, had a coefficient of determination of 0.53 and an MAE of 1.97 Kg.
Cooking requires the usage of various ingredients and cooking utensils. However, if we lack some of them due to insufficient preparation, we cannot cook as expected. In this paper, we propose a cooking support system that complements nonexistent objects with virtual objects. We developed a prototype for cutting apples to evaluate the differences between using the real object and the virtual one. The results showed that the cutting technique equally enhanced both the cases of using a real knife and a virtual knife.
Context-Aware Recommender Systems (CARS) offer context-based suggestions that are particularly crucial in the tourism domain, where personalized experiences significantly enhance user satisfaction. However, the evaluation of CARS is a challenge, partly due to the scarce availability of appropriate datasets that fulfill a variety of evaluation purposes. For example, to evaluate CARS, we need datasets that incorporate context data, but in practice existing datasets provide very little contextual information. This paper presents AUTO-DataGenCARS+, a graphical user-oriented tool designed to generate synthetic data for evaluating both Recommender Systems and CARS. Some of the relevant features of the tool include: a flexible definition of user profiles, user, item and context schemas; a realistic generation of ratings and item attributes; the possibility to mix real and synthetic datasets; functionalities for analyzing and evaluating existing datasets; and an extendable architecture for advanced users. We illustrate the benefits of AUTO-DataGenCARS+ through several examples and experimental evaluations.
This paper presents a method to embed arbitrary bit sequences into acceleration data by attacking capacitive accelerometers with sound. Using a loudspeaker, sound data representing bit sequences is irradiated to the device, manipulating the acceleration data instantly and continuously. Evaluation experiments showed that the extraction accuracy of 10-bit binary information is 100
