International audience

Yoga pose classification is critical for intelligent environments, health, and fitness applications. This study investigates resource-efficient implementations of classification models using contrastive learning frameworks, including SimCLR, MoCo, and BYOL. We evaluate their performance across varying levels of labeled data, focusing on accuracy, computational efficiency, and robustness. MoCo offers a balanced tradeoff with 87.59% accuracy at 50% labeled data, while BYOL achieves strong results with faster inference. SimCLR, suitable for real-time applications due to faster training, consumes more memory and has slower inference. We apply data augmentation and normalization in the preprocessing pipeline to enhance generalization and address challenges like limited data and class imbalance. These techniques improve the model’s resilience and learning efficiency. Our findings guide scalable, energy-efficient, user-centered yoga pose classification models for intelligent environments.

In risk management and investment strategy formulation within financial markets, successfully predicting the Volatility Index (VIX) is crucial. While prediction methods leveraging social media texts have been considered promising, they predominantly rely on Twitter data, leaving the potential of online forums underexplored. However, online forums are rich sources of investor discussions and can provide valuable information for VIX prediction. In this study, we propose a deep learning architecture called ForumPFN that effectively captures the complex discussion structures and contextual dependencies unique to online forums, thereby enhancing financial market predictions. At its core, the Discussion Aggregator Module comprises two main components: a Topic Align Algorithm that classifies and reorganizes posts by topic, and a Multi-Scale 1D-Convolutional Path that integrates features at different scales. This design allows for precise modeling of the discussion flows and dynamics specific to Online Forum, maximizing the utilization of information obtained from online forums. We conduct experiments on directional prediction of the Nikkei 225 Volatility Index (Nikkei 225VI)—a representative VIX of Japanese stocks—using data from Yahoo Finance Message Boards, Japan’s largest online forum. The experimental results confirm that ForumPFN outperforms traditional baseline methods. Furthermore, through ablation studies, we demonstrate the effectiveness of each module in detail and explain the module’s operation via visualization of the attention matrix.

Mental health has a huge impact on humans, affecting both psychological and physical well-being. Excessive stress can lead to depression, reduced productivity, and even suicidal tendencies. Stress also impacts appetite and sleep quality, potentially leading to other health issues. However, stress accumulation often goes unnoticed until it severely impacts health, highlighting the need for daily stress level assessment. This study aims to estimate daily distress levels through natural conversations with a smart speaker. We utilize the audio-visual data of users interacting with a smart speaker on a daily basis, extract features from different modalities through analysis, and predict distress changes in daily life using questionnaire responses as labels. In the experiment, participants interacted with a smart speaker placed in their bedrooms, simulating daily life. Webcam recordings captured facial expressions, voice, and heart rate data, which were preprocessed for analysis. Predictions for happiness, depression, and anxiety levels were made using data from questionnaires filled out after each recording session, with scores ranging from 0 to 18. Results from the 14-day experiment with seven participants, aged 22 to 24, revealed MAEs of 2.04, 2.59, and 2.31 for happiness, depression, and anxiety levels, respectively. The corresponding RMSEs were 2.63, 3.20, and 2.91.

Enabling federated learning in opportunistic networks unlocks the potential for machine learning in challenging environments like disaster zones and remote regions. However, the divergent models induced by dynamic node encounters, combined with complete parameter overlap in model-homogeneous training lead to catastrophic interference, which disrupts training progress. Furthermore, when whole models must be transmitted, nodes with shorter contact duration are limited from participating in the training process. To address these challenges, we propose a different approach for training neural networks in opportunistic settings that leverages independent subnetworks and sequential training. We partition the original neural network into non-overlapping subnetworks and assign each to a unique node. These subnetworks are then trained and exchanged repeatedly during node encounters, exposing them further to diverse datasets. As a consequence, we achieve parallel and conflict-free progress while minimizing participation costs. Our experiments demonstrate that continuous training and subnetwork accumulation foster the development of a more robust model. Moreover, by utilizing pre-trained backbones as feature extractors, we achieve a test accuracy of 75.06% on MEDIC's disaster damage severity assessment task, demonstrating that the approach can be adopted in resource-constrained and dynamic scenarios in the real world.

Regulating human flow is essential to reducing congestion in areas where people gather. A digital twin that realistically simulates human flow helps for this purpose. To realize a realistic human flow simulation mechanism, it is essential to take into account people's attributes. However, existing simulation methods use only location-specific information to predict people's behavior, thus do not reflect the routines that appear in people's actual lives. In this paper, we propose a human flow simulation using synthetic population data that help extract the attributes of people living in a target area. In the proposed method, we simulate the movement of people with each attribute like office workers, students, etc. every 15 minutes using the synthetic population data and the hourly transition probability matrix between PoIs (Points of Interest) by computing the transition probability matrix from hourly PoI-level congestion (people count) in the target area using people trajectory data included in the point-type fluid population data commercially available and applying a Markov chain to the congestion. The proposed simulation mechanism is based on the data assimilation of the actual PoI congestion vector (how many people were staying in each PoI) obtained from the point-type fluid population data and the virtual PoI congestion vector generated from the prediction of people's movement using the attribute information in the synthetic population data at regular time intervals. The data are assimilated at regular intervals to obtain highly accurate PoI-level congestion forecasts. The results of the mobility simulation for office workers showed that the maximum cosine similarity with the actual PoI congestion was 0.96 after 12 hours even when the actual PoI congestion vector is known only for a part of the area (one mesh).

Predicting crowd flow is crucial for decision-making to mitigate various risks. For instance, in social problems such as traffic congestion and over-tourism, countermeasures can be taken by predicting crowd flows in advance. Typically, people visit multiple Points of Interest (PoI) for various purposes. Previous work has proposed methods to incorporate the behavioral characteristics of people in different areas, such as dining areas or office areas, into machine learning models. However, they have not considered the specific behavioral characteristics associated with each PoI, such as when restaurants or train stations experience peak periods. Recently, there has been an increase in the ability to handle large amounts of location information, leading to a growth in the volume of individual trace data. In this study, we propose a crowd flow prediction method that aggregates large-scale individual trace data of movements between PoIs and considers the behavioral characteristics associated with each PoI. We define this information as time series PoI stay counts generated from trace data collected from mobile phones. Using this, we developed a machine learning model to predict the number of people in an area (mesh) over the next several minutes to hours. This prediction is based on the number of people staying at each PoI (category) in neighboring areas (meshes). We applied this approach to densely populated areas in central Tokyo, where congestion is a significant concern, and conducted validation. The results showed that the method utilizing time series PoI stay counts improved prediction accuracy by up to 50% compared to methods that did not use it. Additionally, the Mean Absolute Percentage Error (MAPE) for predicting the number of people staying 1 hour later was only 2.57%.

For reducing investment risks, predicting the volatility of financial markets is crucial. We propose a method for effectively embedding social media posts to facilitate accurate predictions of financial market trends. While discussions on social media inherently consist of paired information - a topic and its sentiment mood - most conventional studies have produced embeddings focusing only on either topic or sentiment information. This approach tends to neglect the intertwined nature of topic and sentiment, thereby overlooking potentially valuable information for market predictions. In this study, we overcome this challenge by introducing a novel document embedding technique that explicitly leverages both topics and sentiments collaboratively for market forecasting. The obtained embeddings are co-trained with financial time series data in a machine learning model, and their efficacy is evaluated through a market prediction task. Benchmarking against various existing document embedding techniques, our model demonstrated superior performance in terms of F-1 Score and Matthews correlation coefficient. The proposed model was further assessed from a practical viewpoint, utilizing investment simulations based on its predictions. These simulations confirmed the model’s potential to generate profits even during heightened market volatility, demonstrating its effectiveness as a real-world investment risk mitigation model. Model interpretation using SHapley Additive exPlanations revealed that while some topic-sentiment pairs on social media consistently contribute to market forecasting, others have only a transient impact. The SHapley Additive exPlanations experiment was also compared to the ablation model and showed that the proposed embedding allows for effective prediction by treating topics and sentiment jointly.

Recently, with the spread of smart houses, the smartness of housing equipment and home appliances has progressed, and the functionality and usability of interfaces between people and equipment and between people and home appliances have become important factors. Currently, the main interfaces are remote controls, smartphone applications, and even voice recognition. Furthermore, research is also being conducted on interfaces that can be operated without having the device at hand using cameras and radio waves. However, special equipment must be installed for operation, and compatibility with room design has become an issue. In this research, we proposed a system to transform existing furniture into an interface rather than providing a new interface. The proposed system focused on vibration sensors that are small, inexpensive, and can be attached to existing furniture or hidden from view. To evaluate the proposed system, an experiment was conducted to transform existing furniture into an interface for swiping by simply attaching the vibration sensor to the existing furniture. Specifically, the system attaches four vibration sensors with synchronized output signals to a table and uses a CNN to learn the vibration data obtained from the sensors to predict the direction of the swipe. As a result, when the table and person swiping were fixed, the system could predict the swipe with an accuracy of over 0.86.

A typical CS1 class involves students working on solving programming problems. Before the pandemic, this occurred in a computer laboratory with a teacher who could quickly assist students having difficulty with their work. Sometimes, there is a need for this intervention even without the student asking for help. An experienced teacher can sense the growing frustration of a student through their overall demeanor. A teacher can also watch how a student codes to provide quick hints to address potential problems. This kind of intervention is challenging to do in an online learning setting. A typical online meeting software provides a small and limited view of a student, often crowded with all the other students. As such, the visual cues of frustration can be easily lost in the noise. Not being able to see the student's code easily is also a problem. The system we are developing aims to create an online IDE that leverages gaze tracking and code logging to automatically identify these struggling students. In the first phase of the research, a learning model will be trained on students' gaze and code logs in line with their overall class performance. The second phase of the research will then use this model to predict the frustration level of student users. Collaboration and gamification strategies will be explored in the final stage of the research that would assist interventions of not just teachers but also classmates who are willing to help.

Abstract To improve the QoL of the elderly, it is essential to predict their stress states. In general, the stress state varies from day to day or time to time depending on what activities are performed and how long/strong. However, most existing studies predict the stress state using biometric data and specific activities (e.g., sleep time, exercise time and amount) as explanatory variables, but do not consider all daily living activities. Therefore, it is necessary to predict the stress state by linking various daily living activities and biometric information. In this paper, we propose a method to improve the prediction accuracy of stress estimation by linking daily living activities data and biometric data. Specifically, we construct a machine learning model in which the objective variable is the result of a stress status questionnaire obtained every morning and evening, and the explanatory variables are the types of daily living activities performed in the 24 h prior to the questionnaire and the feature values calculated from the biometric data during each of the performed activities. The results of the evaluation experiments using the one month data collected from five elderly households, show that the proposed method (using per-activity biometric features) improves the prediction accuracy by more than 10% from the baseline methods (with biometric features without considering activities).

Road traffic censuses have been carried out manually for many years since measurements by machines were not widely spread due to the difficulty of installation. To solve installation difficulty, the size issues of the necessary equipment, and privacy issues of the existing traffic counter, we are conducting research and development of portable traffic counters using a vibration sensor and machine learning. However, vehicle type classification was not realized in the previous work, hence it was not possible to survey traffic volume by vehicle types. In addition, to the best of our knowledge, there is no existing study that can detect and classify vehicles based on road vibrations with a single sensor. In this paper, we propose a method of vehicle type classification that is capable of binary classification of small and large vehicles by machine learning combined with Support Vector Machine and Random Forest for vibrations of passing vehicles. We evaluated the proposed method by conducting measurements for up to 12 hours at two actual road locations. We tested over 5 hours of data and confirmed that small vehicles classified with the F-measure of 0.96 and large vehicles with the F-measure of 0.83.

Systems that detect vehicles passing through roads play a significant role in ITS (Intelligence Transport Systems), due to their wide applicability to traffic monitoring and analysis for road construction/repair planning, congestion, and prediction. Among various systems using cameras, doppler sensors etc., a system that uses road vibration to detect passing vehicles is promising since it has advantages in terms of weather conditions and deployment/operation costs. However, it suffers from the human labor to prepare ground truth labels for training models. In this paper, we propose PAVEMENT, a novel Autonomous Incremental Learning based traffic-census sensor system using a piezoelectric vibration sensor and a video camera without human intervention. PAVEMENT consists of two models: the video-based model which detects vehicles by using bounding boxes (detected by YOLOv3 and DeepSORT) and the vibration-based model which uses road vibrations to detect passing vehicles. To reduce the burden of collecting ground truth labels, we apply linear discriminant analysis and incremental learning to train the vibration-based model by using the result of the video-based model as ground truth. Once the vibration-based model is trained, it can be used for traffic census on roads without the video camera for various conditions (weather, lighting, and other environmental factors). We collected the video and vibration data of more than 4,000 passing vehicles on roads in different places and applied our method to the data. As a result, PAVEMENT achieved over 98.4% accuracy and 98.0% f1-score in detecting passing vehicles using the model trained with 15 incremental learning steps in 1 minute interval.

In recent years, the use of energy harvesting (EH) sensors has led to the proposal of activity sensing systems that are easy to install and do not require maintenance such as battery replacement. In this study, we aim to construct a system that can not only sense but also recognize activities of daily living (ADLs) using only power generated by EH sensors. To achieve this goal, in this paper, we propose a fully EH-based ADL recognition system called Batterfly, which consists of EH analog PIR sensor nodes that can operate with indoor light, continuously senses human movement, and recognizes daily activities through machine learning. We applied the distributed execution method of the activity recognition model with five sensor nodes to five types of activities by five participants, and found that the system could recognize them with an average accuracy of 63.59%, comparable to the performance of the centralized model running on a gateway.

In the tourism sector, user-generated information and communication among tourists are perceived to be more effective and reliable contents. In addition, the collection of dynamic tourism information with high spatio-temporal resolution is required to provide comfortable tourism in response to the changing tourism style with the advancement of information technology. Participatory sensing, which can collect various types of information, is a useful method by which to collect these contents. However, continuous participation of users is essential in participatory sensing, and it is one of the most important points to stimulate participation motivation. In the tourism situation, we also need to pay attention to the total tourist satisfaction of participants. In this study, we investigate the effects of task allocation interfaces and user types on the efficiency of tourism information collection, tourism behavior and satisfaction in gamified participatory sensing. Two types of task allocation interfaces (free selection and agent interaction) were designed and implemented, and a sightseeing experiment was conducted with 10 participants at an actual sightseeing spot (Nara, Japan). As a result, we found that there was no difference in the effect of each interface on sightseeing satisfaction, but the characteristics of the collected data that, free selection allows for the collection of quantitative data and agent interaction allows for the efficient collection of data needed by the system, were different. In addition, we found that different user types had different tendencies for their contribution to sensing and their interface preferences.

Street parking in prohibited areas has become a social problem, especially in urban and tourist areas. In addition, because street parking can cause traffic congestion and accidents, real-time detection is required. The detection of street parking has been previously implemented on the basis of comparisons of videos recorded by fixed-point cameras. However, this approach has a limited detection area and low accuracy. In this demonstration, we present a system that recognizes street parking in real-time using a model trained by dashboard camera videos, which are widely used. The trained model was constructed by collecting data on 1,765 vehicles from dashboard camera videos. We use the Jetson TX2 as an edge device for vehicle recognition and processing. We create and demonstrate a dashboard camera device by attaching a camera and sensor module to the Jetson TX2.

In recent years, there has been a growing interest in travel applications that provide on-site personalized tourist spot recommendations. While they are generally useful, most of the available apps are focused on helping tourists make decisions only on the next spot to visit. This may cause that the tourists miss attractive spots to visit in the future. Due to the lack of awareness on the spots to go afterwards, they are unable to visit the spots they wanted to visit later, hence their overall tourism satisfaction decreases. In this study, we introduce an on-site tourism recommendation system, ISO-Tour, which can be used on the spot during the tour and allows users to consider multiple spots to visit next taking into account the trade-off between satisfaction of the next spot and that of the subsequent spots visited in the future.

In this paper, we aim to reveal the relationship between the number of people infected with COVID-19 in each ward in Tokyo and the changes in human mobility behavior using demographic information (population density, number of restaurants, etc.) and mobility data collected from GPS data of residents in Tokyo's 23 wards. The results confirmed that changes in human mobility behavior extracted from mobility data in each ward were an important feature related to the number of people infected by COVID-19 on the previous day's difference. These results suggest that the transition of the number of infected people in COVID-19 is largely due to human mobility behavior.

As an increasing demand for human activity monitoring in many smart services such as elderly monitoring, there is a keen need of an activities of daily living (ADL) recognition system. which can be easily deployed in ordinary homes and does not require periodic maintenance such as battery replacement for long time. In this paper, we propose an ADL recognition system which can run continuously without feeding power from outlets by intermittent sensing and distributed processing of energy harvesting (EH) sensor modules. Specifically, we have designed and developed an EH sensor node composed of (i) a micro-controller board with an analog PIR sensor which senses human activity as analog signals and form a BLE mesh network with other sensor nodes and (ii) an energy harvest module with solar panels and a rechargeable battery. We have also implemented a simple distributed random forest (RF) classifier consisting of multiple RF classifiers trained independently and running on different nodes which exchange the classification results with each other via BLE and make a final decision based on majority vote. Through experiments with five sensor nodes deployed in our smart home testbed, the distributed RF classifier classified the collected data of up to five different activities with average accuracy of over 90%.

A wearable based long-term lifelogging system is desirable for the purpose of reviewing and improving users lifestyle habits. Energy harvesting (EH) is a promising means for realizing sustainable lifelogging. However, present EH technologies suffer from instability of the generated electricity caused by changes of environment, e.g., the output of a solar cell varies based on its material, light intensity, and light wavelength. In this paper, we leverage this instability of EH technologies for other purposes, in addition to its use as an energy source. Specifically, we propose to determine the variation of generated electricity as a sensor for recognizing "places" where the user visits, which is important information in the lifelogging system. First, we investigate the amount of generated electricity of selected energy harvesting elements in various environments. Second, we design a system called EHAAS (Energy Harvesters As A Sensor) where energy harvesting elements are used as a sensor. With EHAAS, we propose a place recognition method based on machine-learning and implement a prototype wearable system. Our prototype evaluation confirms that EHAAS achieves a place recognition accuracy of 88.5% F-value for nine different indoor and outdoor places. This result is better than the results of existing sensors (3-axis accelerometer and brightness). We also clarify that only two types of solar cells are required for recognizing a place with 86.2% accuracy.

We developed a simulation program for trading Nikkei stock index options and verifies the validity of the volatility index (VIX) prediction model proposed by Suwa et al. (2017). We simulated two cases from 18 Nov. 2014 to 29 Jun. 2016. One case involved a benchmark of trading every day during that period and the other was in accordance with the buy/sell/hold instructions of Suwa et al.'s VIX prediction model. When using the call option butterfly spread according to their model's instructions, profit increased from -3,926 to 536 yen. When using the put option butterfly spread according to their model's instructions, profit increased from -4,818 to -799 yen. Therefore, Suwa et al.'s VIX prediction model is effective.

In an era where billions of IoT devices are deployed, edge/fog computing paradigms are attracting attention for their ability to reduce processing delays and mitigate waste of communication resources. However, since the computing system assumed by edge/fog paradigms have heterogeneity (in terms of the computing power of devices, network performance between devices, device density, etc.), provisioning computational resources according to computational demand becomes a challenging constrained optimization problem. In this paper, we propose in-situ resource provisioning method consisting of insitu resource area selection with adaptive scale out and in-situ task scheduling based on tabu search algorithm. We conducted a simulation study in a target regional area where 2,000 IoT devices and 10 IoT services are deployed to evaluate the effectiveness of the proposed algorithm. The simulation results show that our proposed algorithm can obtain higher user QoS compared to conventional resource provisioning algorithms.

These days, smart home applications such as a concierge service for residents, home appliance control and so on are attracting attention. In order to realize these applications, we strongly believe that we need a system which recognizes the various human activities accurately with a low cost device. There are many studies which work on the activity recognition in the smarthome. Moreover, we also have proposed the activity recognition technique in the smarthome by utilizing the digitaloutput-PIR sensor, door sensor, watt meter. However, the study has the challenge: we cannot distinguish between the similar tiny activities at the same place: “eating” and “reading” with sitting on a sofa. In order to cope with this challenge, we introduce ALPAS: analog-output-PIR-sensor-based activity recognition technique which recognizes the detailed activities of the user. Our technique recognizes the activity of the user by utilizing the machine learning. We evaluated the proposed technique in a smarthome which belongs to the authors' university. In the evaluation, three subjects performed four different activities with sitting on a sofa. As a result, we achieved F-Measure: 57.0%.

Recently, as elderly people population grows, the burdens on caretakers are getting larger. In daycare centers, caretakers make a daycare report aiming to improve the senior citizen's Quality of Life. However, in the present situation, it is difficult for caretakers to record the senior citizen's activity in detail, since each caretaker needs to take care of several senior citizens at the same time. To reduce the burden of caretakers, many elderly monitoring systems have been proposed so far, but most of them are not effective in the sense that they force the senior citizen to use dedicated devices such as smart phone and/or particular applications that are obtrusive and cumbersome for care receivers. In this paper, we propose a semi-automatic care-taking report generation system which can monitor movements/activity of senior citizens in daycare centers. Our proposed system estimates multiple locations (areas) where senior citizens are located with the BLE beacon, by utilizing RSSI of the Bluetooth radio wave. Also, the accelerometer implemented in the tag estimates the activity of the elderly. The information of the estimated area and activity is stored in a server with time stamp. The server generates the daycare report based on it. In order to evaluate the proposed system, we have deployed our system in a daycare center: Ikoi-no-ie 26. Evaluation result in Ikoi-no-ie 26 showed that our system estimated the subject's present area with F-measure: 80.6% and activity with F-measure: 73.8% and generated the daycare report.

We propose a method of predicting an increase in the Nikkei VI index by analyzing social media based on the premise that investor sentiment is posted on social media. Since the VI index expresses the fear of investors, it is a closely related index to the risk of depression. Therefore, the VI index is an important indicator as an instrument for investment judgment. To predict the increase in the VI index more accurately, we divide messages by topic models specific to social media of stock trading and predict such the increase by machine learning using those topics. As a result of leave-one-day-out cross-validation, precision of our method was 0.45. We also found that the daily fluctuation in the VI index and the number of messages are as effective as feature quantities as the topic-posting frequency.

A key goal of this research is to understand the relationship between adoption of software library versions and its release cycle. In detail, we conducted an empirical study of the release cycle of 23 libraries and how they were adopted by 415 Apache Software Foundation (ASF) client projects. Our preliminary findings show that software projects are quicker to update earlier rapid-release libraries compared to library projects with a longer release cycle.

EdgeComputing and Fog Computing are new paradigms where data processing is executed in or on the edge of networks to mitigate cloud server load. However, EdgeComputing and Fog Computing still need powerful servers on the edge of networks which impose additional costs for deployments. We proposed a platform called IFoT (Information Flow of Things) that efficiently performs distributed processing as well as distribution and analysis of data streams near their sources based on "Process On Our Own (PO3)" concept. In IFoT, processing of tasks for cloud servers is delegated to an ad-hoc distributed system consisting of proximity IoT devices for distributed real-time stream processing. In this demonstration, we show a face recognition system for person tracking developed on top of IFoT middleware which locally processes video streams in real-time and in a distributed manner by using computational resources of IoT devices.

Family Tracing and Reunification (FTR) is the process whereby families separated by disasters are reunited. Current FTR systems use either inefficient paper-based forms and notice boards or digital registries that need the Internet, which may be unavailable during disasters. In this demonstration we present Milk Carton: a system that aids in FTR. Milk Carton creates a registry containing evacuee records. To find separated persons, Milk Carton uses Eigenfaces face recognition to match queries with existing records. Milk Carton uses a clustered architecture of Computing Nodes to handle data storage and execute the Eigenfaces algorithm. To operate under challenged-network environments, Milk Carton uses response patrol vehicles as data ferries to deliver data. In this demonstration, we show how Milk Carton uses Eigenfaces to locate separated persons and how data ferries and Computing Nodes function.

These days, smart home applications such as a concierge service for residents, home appliance control and so on are attracting attention. In order to realize these applications, we strongly believe that we need an indoor positioning system which fulfills the following requirements: Req 1: high accuracy; Req 2: low installation cost; Req 3: small burden on the user; Req 4: low privacy invasion. There are several studies which work on the indoor positioning system. However, these previous works do not accomplish the requirements. In this paper, we present a piezo sensor-based indoor positioning system which estimates the position of the user by utilizing a piezo component attached on the floor. To realize the proposed positioning system, we have tackled two challenges. First challenge is the development of an indoor positioning technique. We cannot utilize TDoA technique that is used to estimate the distance from the target, since the calculation of vibration velocity is difficult. To cope with this challenge, we have developed a new technique which estimates the position of the user from floor vibrations caused by their actions. Second challenge is the selection of the feature vector to estimate the vibration type accurately. We have selected MFCC, FFT, and Envelope shape features from preliminary experiments. We have implemented the proposed system in our smart home testbed. We have evaluated the performance of the vibration type estimation technique. As a result, we have confirmed that our technique estimates the type with F-measure: 93.9%.

Recently, as elderly people population grows, the burdens on caretakers are getting larger. In daycare center, caretakers make a daycare report aiming to improve the senior citizen's Quality of Life. However, in the aged society, it is difficult for caretakers to record the senior citizen's activity in detail, since each caretaker needs to take care of several senior citizens at the same time. In this paper, we propose a semi-automatic day-care report generation system. Our proposed system estimates the present area and activity of the senior citizen by utilizing the accelerometer-implemented-BLE-beacon tags and generate the daycare report. Based on the generated report, the caretaker works effectively. Evaluation result in a daycare center showed that our system estimated the subject's present area with F-measure: 80.6% and activity with F-measure: 73.8% and automatically generated the daycare report.

In this paper, we aim to develop a method to create personalized and high-presence multi-channel contents for a sport game through realtime content curation from various media streams captured/created by spectators. We use the live TV broadcast as a ground truth data and construct a machine learning-based model to automatically conduct curation from multiple videos which spectators captured from different angles and zoom levels. The live TV broadcast of a baseball game has some curation rules which select a specific angle camera for some specific scenes (e.g., a pitcher throwing a ball). As inputs for constructing a model, we use meta data such as image feature data (e.g., a pitcher is on the screen) in each fixed interval of baseball videos and game progress data (e.g., the inning number and the batting order). Output is the camera ID (among multiple cameras of spectators) at each point of time. For evaluation, we targeted Spring-Selection high-school baseball games. As training data, we used image features, game progress data, and the camera position at each point of time in the TV broadcast. We used videos of a baseball game captured from 7 different points in Hanshin Koshien Stadium with handy video cameras and generated sample data set by dividing the videos to fixed interval segments. We divided the sample data set into the training data set and the test data set and evaluated our method through two validation methods: (1) 10-fold crossvalidation method and (2) hold-out methods (e.g., learning first and second innings and testing third inning). As a result, our method predicted the camera switching timings with accuracy (F-measure) of 72.53% on weighted average for the base camera work and 92.1% for the fixed camera work.

We analyze the relations between the stock market and a stock bulletin board system (BBS) in Japan. Previous studies in the USA found that the characteristics of messages posted on stock BBSs can predict market volatility and trading volume. We develop hypotheses based on the results of those analyses and apply statistical analysis to the data about companies mentioned in a large number of messages posted on the Yahoo! stock message board in Japan in 2005-2006. We analyze the contents of these messages using natural language processing. We find a significant correlation between the number of postings and market volatility and trading volume, and also find significant correlation between the amount of bullish and bearish opinion and the stock return.