Inspired by the progress of motion synthesis models, we leverage cross-modality transfer to generate realistic synthetic Inertial Measurement Unit (IMU) data from textual descriptions, hence Text2IMU. We use an established motion synthesis model and textual descriptions to generate sequences of 3D human activities. To obtain realistic and diverse sensor readings, we created multiple body surface models with different body morphologies. With the text prompts, we let the surface models perform activities and synthesise acceleration and gyroscope data for multiple virtual IMU positions. We show that synthetic data, generated by Text2IMU, can be used to classify activities across three public benchmark datasets. We demonstrate that our Text2IMU synthesis approach does not require measured data of the target domain. Text2IMU yields an average Human Activity Recognition (HAR) accuracy of$\text{7 9. 2 \%}$for correctly synthesised activities, which doubles the performance of synthetic sensor data obtained from baseline models. We demonstrate that synthetic HAR model training can replace empirical data acquisition when the prompted activities can be successfully generated.

We analyse pretrained and non-pretrained deep neural models to detect 10-seconds Bowel Sounds (BS) audio segments in continuous audio data streams. The models include MobileNet, EfficientNet, and Distilled Transformer architectures. Models were initially trained on AudioSet and then transferred and evaluated on 84 hours of labelled audio data of eighteen healthy participants. Evaluation data was recorded in a semi-naturalistic daytime setting including movement and background noise using a smart shirt with embedded microphones. The collected dataset was annotated for individual BS events by two independent raters with substantial agreement (Cohen's Kappa κ = 0.74). Leave-One-Participant-Out cross-validation for detecting 10-second BS audio segments, i.e. segment-based BS spotting, yielded a best F1 score of 73% and 67%, with and without transfer learning respectively. The best model for segment-based BS spotting was EfficientNet-B2 with an attention module. Our results show that pretrained models could improve F1 score up to 26%, in particular, increasing robustness against background noise. Our segment-based BS spotting approach reduces the amount of audio data to be reviewed by experts from 84 h to 11 h, thus by ∼ 87%.

We analyse the influence of reflective photoplethysmography (PPG) sensor positioning relative to blood vessels. A voxel based Monte Carlo simulation framework was developed and validated to simulate photon-tissue interactions. An anatomical model comprising a multi-layer skin description with a blood vessel is presented to simulate PPG sensor positioning at the volar wrist. The simulation framework was validated against standard test cases reported in literature. The blood vessel was considered in regular and dilated states. Simulations were performed with 108photon packets and repeated five times for each condition, including wavelength, relative position of PPG sensor and vessel, and vessel dilation state. Statistical weights were associated to photon packets to represent absorption and scattering effects. A symmetrical arrangement of the PPG sensor around the blood vessel showed the maximum AC signal. When the PPG sensor was not centrally placed over the vessel, simulated photon weight in systolic and diastolic state deteriorated by ≥5% for both wavelengths. With a position-dependent variation of ≥5% at 660 nm and ≥12% at 940 nm of light absorption, blood had the most profound effect on signal quality. The mean penetration depth is dependent on the blood vessel position for both wavelengths. Our simulation results demonstrate the susceptibility of reflective PPG measurement to interference and could explain wearable PPG sensor performance variations related to positioning and wavelength.

We investigate non-contact sensing of temporalis muscle contraction in smart eyeglasses frames to detect eating activity. Our approach is based on infra-red proximity sensors that were integrated into sleek eyeglasses frame temples. The proximity sensors capture distance variations between frame temple and skin at the frontal, hair-free section of the temporal head region. To analyse distance variations during chewing and other activities, we initially perform an in-lab study, where proximity signals and Electromyography (EMG) readings were simultaneously recorded while eating foods with varying texture and hardness. Subsequently, we performed a free-living study with 15 participants wearing integrated, fully functional 3Dprinted eyeglasses frames, including proximity sensors, processing, storage, and battery, for an average recording duration of 8.3hours per participant. We propose a new chewing sequence and eating event detection method to process proximity signals. Free-living retrieval performance ranged between the precision of 0.83 and 0.68, and recall of 0.93 and 0.90, for personalised and general detection models, respectively. We conclude that noncontact proximity-based estimation of chewing sequences and eating integrated into eyeglasses frames is a highly promising tool for automated dietary monitoring. While personalised models can improve performance, already general models can be practically useful to minimise manual food journalling.

We present a modelling and simulation framework to synthesise body-worn inertial sensor data based on personalised human body surface and biomechanical models. Anthropometric data and reference images were used to create personalised body surface mesh models. The mesh armature was aligned using motion capture reference pose and afterwards mesh and armature were parented. In addition, skeletal models were created using an established musculoskeletal dynamic modelling framework. Four activities of daily living (ADL), including upper and lower limbs were simulated with surface and skeletal models using motion capture data as stimuli. Acceleration and angular velocity data were simulated for 12 body areas of surface models and 8 body areas of skeletal models. We compared simulated inertial sensor data of both models against physical IMU measurements that were obtained simultaneously with video motion capture. Results showed average errors of 27 °/s vs. 31 °/s and 1.7 m/s2vs. 3.3 m/s2for surface and skeletal models, respectively. Mean correlation coefficients of body surface models ranged between 0.2 – 0.9 for simulated angular velocity and between 0.1 – 0.8 for simulated acceleration when compared to physical IMU data. The proposed surface modelling consistently showed similar or lower error compared to established skeletal modelling across ADLs and study participants. Body surface models can offer a more realistic representation compared to skeletal models for simulation-based analysis and optimisation of wearable inertial sensor systems.

This paper investigates a context-adaptive sample acquisition strategy at sub-Nyquist sampling rate for wearable embedded sensor devices. Our approach can be applied to compressive sensing frameworks to minimise sampling and transmission costs. We consider a context estimate to represent the local signal structure and a feed-forward response model to continuously tune signal acquisition of an online sampling and transmission system. To evaluate our approach, we analysed the performance in different pattern recognition scenarios. We report three case studies here: (1) eating monitoring based on electromyography measurements in smart eyeglasses, (2) human activity recognition based on waist-worn inertial sensor data, and (3) heartbeat detection and arrhythmia classification based on single-lead electrocardiogram readings. Compared to conventional sub-Nyquist sampling, our context-adaptive approach saves between 13 to 22 percent of energy, while achieving similar pattern recognition performance and reconstruction error.

We propose a framework to automatically extract body landmarks and related measurements from 3D body scans and replace manual body shape estimation in fitting smart garments. Our framework comprises five steps: 3D scan acquisition and segmentation, 2D image conversion, extraction of body landmarks using a Convolutional Neural Network (CNN), back projection and mapping of extracted landmarks to 3D space, body measurements estimation and tailored garment generation. We trained and tested the algorithm on 3000 synthetic 3D body models and estimated body landmarks required for T-Shirt design. The results show that the algorithm can successfully extract 3D body landmarks of the upper front with a mean error of 1.01 cm and of the upper back with a mean error of 0.78 cm. We validated the framework the framework in automated tailoring of an electrocardiogram (ECG)-monitoring shirt based on the predicted landmarks. The ECG shirt can fit all evaluated body shapes with an average electrode-skin distance of 0.61 cm.

The papers in this special section focus on cloud computing, fog computing, the Internet of Things, and Big Data analytics for the future healthcare industry, or Healthcare 4.0. Healthcare Industry 4.0 allows increasing flexibility in production, speeding up both manufacturing and market processes, increasing both the product quality and productivity, and changing business models modifying the interaction with value chain, competitors, and clients. Healthcare Industry 4.0 requires investments and mind-set change for cross-industry collaboration, agreements on data ownership, security, legal issue solving, product registration standards, new machine-to-machine communication protocols, and employment/skills development. Furthermore,Healthcare Industry 4.0 is revolutionizing the market of health service provisioning to patients and clinical operators.

We investigate an approach to analyse and represent the functional range of motion (fROM) in patients after stroke during long-term rehabilitation in a day-care centre. Movement data of patients after stroke were recorded during free activities of living (ADL) and one-to-one (OTO) therapies using wearable inertial motion sensors. Using a gradient descent sensor fusion algorithm, we estimated orientation of upper body extremities and described extremity positions as frequency statistics in spatial posture cubics. We visualised the fROM and compared sensor-based posture representation with video reference during OTO. To illustrate our approach, we analysed differences in affected and non-affected arm use in three typical patients after stroke across multiple weeks. Our analysis revealed differences in body sides as well as between ADL and OTO. Posture cubics may provide clinicians with an intuitive tool for longitudinal fROM analysis.

This paper presents a motion-adaptive approach to duty-cycling the orientation estimation of a wearable inertial measurement unit (IMU). Specifically, a proportional forward-controller was employed to dynamically tune the sampling and orientation update rate of a Madgwick filter. An energy model was defined to estimate the power consumed by individual inertial sensors and processing elements. We demonstrate the efficacy of our controller by analysing multi-day free-living motion recordings of wrist-worn IMUs. In a comparison of the orientation estimation between the full duty-cycle and the adaptive one, average error was approx. 10 degrees while saving more than 30% of sensor node energy. To assess the orientation information retained by our approach, we analysed the binary pattern classification performance for recognising food and fluid intake gestures. Recognition performance of the motion-adaptive duty-cycling remained above 90% up to an energy saving of approx. 32.5%, confirming the profound potential of the method.

We present a spotting network composed of Gaussian Mixture Hidden Markov Models (GMM-HMMs) to detect sparse natural gestures in free living. The key technical features of our approach are (1) a method to mine non-gesture patterns that deals with the arbitrary data (Null Class), and (2) an optimisation based on multipopulation genetic programming to approximate spotting network's parameters across target and non-target models. We evaluate our GMM-HMMs spotting network in a novel free living dataset, including totally 35 days of annotated inertial sensor's recordings from seven participants. Drinking was chosen as target gesture. Our method reached an average F1-score of over 74% and clearly outperformed an HMM-based threshold model approach. The results suggest that our spotting network approach is viable for sparse natural pattern spotting.

We design and 3D print conductive lines and EMG electrodes on eyeglasses temples. We evaluate the electrical property and the EMG signal quality of the printed components and report line resistance, electrode surface resistance, and EMG signal quality. We found that the signal quality is comparable to non-printed lines and electrodes. Our work shows that 3D printing of conductive lines and electrodes on custom-shaped eyeglasses frames is feasible for chewing monitoring.

We propose a mistake-driven skill estimation approach for movement analysis in sports using wearable inertial measurement units (IMUs) and continuous regression models. As a motion-quality oriented sport, we focus on Nordic Walking. Nordic Walking involves complex body part interactions, which when wrongly performed, increase injury risk and reduce training effectiveness. We present an approach to assess three mistakes related to health risk which are typical in beginners. Based on a stride segmentation, features relevant to the mistakes were extracted and selected in a dedicated feature selection step. Subsequently, Bayesian Ridge Regression (BRR) and alternative regression models were trained for each mistake type. Models were evaluated in our pattern analysis architecture supporting parallel and continuous step-wise estimation of all mistakes in movement skill grades ranging from 1 (correct) to 3 (incorrect). We evaluated our skill estimation approach in a study including 10 Nordic Walking beginners and 11247 expert-annotated strides derived from 50 recording sessions. We investigate seven practical wearable sensor placement configuration using leave-one-participant-out (LOPO) cross-validation. Results showed that all mistakes can be estimated with an normalised RMSE of 24.15 % across all participants. Additional analysis of trends suggest that participants could improve skilfulness in training sessions during our study.

We propose to 3-D-print personal fitted regular-look smart eyeglasses frames equipped with bilateral electromyography recording to monitor temporalis muscles' activity for automatic dietary monitoring. Personal fitting supported electrode-skin contacts are at temple ear bend and temple end positions. We evaluated the smart monitoring eyeglasses during in-lab and free-living studies of food chewing and eating event detection with ten participants. The in-lab study was designed to explore three natural food hardness levels and determine parameters of an energy-based chewing cycle detection. Our free-living study investigated whether chewing monitoring and eating event detection using smart eyeglasses is feasible in free-living. An eating event detection algorithm was developed to determine intake activities based on the estimated chewing rate. Results showed an average food hardness classification accuracy of 94% and chewing cycle detection precision and recall above 90% for the in-lab study and above 77% for the free-living study covering 122 hours of recordings. Eating detection revealed the 44 eating events with an average accuracy above 95%. We conclude that smart eyeglasses are suitable for monitoring chewing and eating events in free-living and even could provide further insights into the wearer's natural chewing patterns.

We present and evaluate measurement fusion and decision fusion for recognizing apnea and periodic limb movement in sleep episodes. We used an in-bed sensor system composed of an array of strain gauges to detect pressure changes corresponding to respiration and body movement. The sensor system was placed under the bed mattress during sleep and continuously recorded pressure changes. We evaluated both fusion frameworks in a study with nine adult participants that had mixed occurrences of normal sleep, apnea, and periodic limb movement. Both frameworks yielded similar recognition accuracies of 72.1 ± ∼ 12% compared to 63.7 ± 17.4% for a rule-based detection reported in the literature. We concluded that the pattern recognition methods can outperform previous rule-based detection methods for classifying disordered breathing and period limb movements simultaneously.

We utilise smart eyeglasses for dietary monitoring, in particular to sense food chewing. Our approach is based on a 3D-printed regular eyeglasses design that could accommodate processing electronics and Electromyography (EMG) electrodes. Electrode positioning was analysed and an optimal electrode placement at the temples was identified. We further compared gel and dry fabric electrodes. For the subsequent analysis, fabric electrodes were attached to the eyeglasses frame. The eyeglasses were used in a data recording study with eight participants eating different foods. Two chewing cycle detection methods and two food classification algorithms were compared. Detection rates for individual chewing cycles reached a precision and recall of 80%. For five foods, classification accuracy for individual chewing cycles varied between 43% and 71%. Majority voting across intake sequences improved accuracy, ranging between 63% and 84%. We concluded that EMG-based chewing analysis using smart eyeglasses can contribute essential chewing structure information to dietary monitoring systems, while the eyeglasses remain inconspicuous and thus could be continuously used.

In this paper, we present a method to estimate oxygen uptake ( VO2) during daily life activities and transitions between them. First, we automatically locate transitions between activities and periods of nonsteady-state VO2. Subsequently, we propose and compare activity-specific linear functions to model steady-state activities and transition-specific nonlinear functions to model nonsteady-state activities and transitions. We evaluate our approach in study data from 22 participants that wore a combined accelerometer and heart rate sensor while performing a wide range of activities (clustered into lying, sedentary, dynamic/household, walking, biking and running), including many transitions between intensities, thus resulting in nonsteady-state VO2. Indirect calorimetry was used in parallel to obtain VO2 reference. VO2 estimation error during transitions between sedentary, household and walking activities could be reduced by 16% on average using the proposed approach, compared to state of the art methods.

We present a framework to mine relations and group variables that represent measurement and status information from sensors and actuators in office buildings. Our work is motivated by the need to manage growing numbers of devices and related automation functions in buildings that are currently often manually commissioned and maintained. Our approach relies on the idea that building variables at the same location will change value in a temporal relation that can be discovered. Based on event sequences derived from various variables and modalities, our approach initially mines temporal association rules and subsequently groups variables. We propose a weighted transitive clustering (WTC) algorithm to automatically group co-located building variables. To validate our approach, we used living-lab office recordings across 14 months from three different office rooms. We compare our approach against a random guess baseline, a hierarchical agglomerative clustering (HAC) approach, and the rules of a manually configured building management system (BMS). We found that within three months of operation, 75% of the building variables could be grouped. Our WTC approach outperforms the baseline and HAC. Furthermore, we show that our framework can be used develop different BMS applications, including counting people in building spaces and identifying BMS configuration errors.

We introduce a hierarchical nonparametric topic modeling approach to infer activity routines from context sensor data streams based on a distance dependent Chinese restaurant process (ddCRP). Our approach does not require labeled data at any stage. Neither does our approach depend on time-invariant sliding windows to sample context word statistics. Our activity discovery approach builds on the idea that context words occurring within one activity are semantically similar, whereas context words of different activities are less similar. Context word streams are segmented into supersamples and then semantic and temporal features are obtained to construct a segmentation prior that relates supersamples via its context words. Our hierarchical model uses the segmentation prior and ddCRP to group supersamples and the Chinese restaurant process (CRP) to discover activities. We evaluate our approach using the Opportunity dataset that contains activities of daily living. Besides being nonparametric, our ddCRP based model outperforms both, classic parametric latent Dirichlet allocation (LDA) and the nonparametric Chinese restaurant franchise (CRF). We conclude that ddCRP+CRP is an adequate approach for fully unsupervised activity discovery from context sensor data.

We present a novel sensor system for the support of nutrition monitoring. The system is based on smart table cloth equipped with a fine grained pressure textile matrix and a weight sensitive tablet. Unlike many other nutrition monitoring approaches, our system is unobtrusive, non privacy invasive and easily deployable in every day life. It allows the spotting and recognition of food intake related actions, such as cutting, scooping, stirring, etc., the identification of the plate/container on which the action is executed, and the tracking of the weight change in the containers. In other words, we can determine how many pieces are cut on the main dish plate, how many are taken from the side dish, how many sips are taken from the drink, how fast the food is being consumed and how much weight is taken overall. In addition, the distinction between different eating actions, such as cutting, scooping, poking, provides clues to the type of food taken and the way the meal is consumed. We have evaluated our system on 40 meals (5 subjects) in a real life living environment: for seven eating related actions (cutting, scooping, stirring, etc.), resulting in above 90% average recognition rate for person dependent cases, and spotting each action out of continuous data streams (average F1 score 87%).

We investigate the process of transferring the activity recognition models within the nodes of a body sensor network (BSN). In particular, we propose a methodology that supports and makes the transferring possible. Based on a collaborative training strategy, classifier ensembles of randomized trees are used to create activity recognition models that can successfully be transferred within the nodes of the network. The methodology has been applied in scenarios where a node present in the network is replaced by a new node located in the same position (replacement scenario) and relocated to a previously unknown position (relocation scenario). Experimental results show that the transferred recognition models achieve high-recognition performance in the replacement scenario and good-recognition performance are achieved in the relocation scenario. Results have been validated with multiple K-folds cross-validations in order to test the performance of the methodology when different amount of data are shared between nodes.

We introduce an approach to personalize energy expenditure (EE) estimates in free living. First, we use topic models to discover activity composites from recognized activity primitives and stay regions in daily living data. Subsequently, we determine activity composites that are relevant to contextualize heart rate (HR). Activity composites were ranked and analyzed to optimize the correlation to HR normalization parameters. Finally, individual-specific HR normalization parameters were used to normalize HR. Normalized HR was then included in activity-specific regression models to estimate EE. Our HR normalization minimizes the effect of individual fitness differences from entering in EE regression models. By estimating HR normalization parameters in free living, our approach avoids dedicated individual calibration or laboratory tests. In a combined free-living and laboratory study dataset, including 34 healthy volunteers, we show that HR normalization in 14-day free-living data improves accuracy compared to no normalization and normalization based on activity primitives only ( 29.4% and 19.8 % error reduction against lab reference). Based on acceleration and HR, both recorded from a necklace, and GPS acquired from a smartphone, EE estimation error was reduced by 10.7 % in a leave-one-participant-out analysis.

Several methods to estimate energy expenditure (EE) using body-worn sensors exist; however, quantifications of the differences in estimation error are missing. In this paper, we compare three prevalent EE estimation methods and five body locations to provide a basis for selecting among methods, sensors number, and positioning. We considered 1) counts-based estimation methods, 2) activity-specific estimation methods using METs lookup, and 3) activity-specific estimation methods using accelerometer features. The latter two estimation methods utilize subsequent activity classification and EE estimation steps. Furthermore, we analyzed accelerometer sensors number and on-body positioning to derive optimal EE estimation results during various daily activities. To evaluate our approach, we implemented a study with 15 participants that wore five accelerometer sensors while performing a wide range of sedentary, household, lifestyle, and gym activities at different intensities. Indirect calorimetry was used in parallel to obtain EE reference data. Results show that activity-specific estimation methods using accelerometer features can outperform counts-based methods by 88% and activity-specific methods using METs lookup for active clusters by 23%. No differences were found between activity-specific methods using METs lookup and using accelerometer features for sedentary clusters. For activity-specific estimation methods using accelerometer features, differences in EE estimation error between the best combinations of each number of sensors (1 to 5), analyzed with repeated measures ANOVA, were not significant. Thus, we conclude that choosing the best performing single sensor does not reduce EE estimation accuracy compared to a five sensors system and can reliably be used. However, EE estimation errors can increase up to 80% if a nonoptimal sensor location is chosen.

Ubiquitous body-worn sensors and mobile devices have extensive applications in health monitoring and user guidance when combined with advanced algorithms. This special issue includes 13 papers, presenting novel approaches and trends in both sensor integration and data processing algorithms. It features some of the best papers from the International Conference on Wearable and Implantable Body Sensor Networks (BSN) in 2013 and 2014, as well as normal submissions to J-BHI on this topic. The scope of this paper spans from motion analysis to cardiac and respiratory monitoring, and from Electroencephalogram and Electromyogram processing to core body temperature estimation and radio propagation.

In this work we investigate the process of transferringthe activity recognition models of the nodes of a BodySensor Network and we proposed a methodology that supportsand makes the transferring possible. The methodology, based on acollaborative training strategy, makes use of classifier ensemblesof randomised trees that allow to generate activity recognitionmodels able to be successfully transferred through the nodes ofthe network. Experimental results evaluated on 17 subjects witha network of 5 wearable nodes with 5 everyday life activities showthat the recognition models can be transferred to a new untrainednode replacing a node previously present in the network withouta significant loss in the recognition performance. Moreover, themodels achieve good recognition performance in nodes located inpreviously unknown positions.

Body sensor networks (BSNs) have provided the opportunity to monitor energy expenditure (EE) in daily life and with that information help reduce sedentary behavior and ultimately improve human health. Current approaches for EE estimation using BSNs require tedious annotation of activity types and multiple body sensor nodes during data collection and high accuracy activity classifiers during post processing. These drawbacks impede deploying this technology in daily life — the primary motivation of using BSNs to monitor EE. With the goal of achieving the highest EE estimation accuracy with the least invasiveness and data collection effort, this paper presents an unsupervised, single-node solution for data collection and activity clustering. Motivated by a previous finding that clusters of similar activities tend to have similar regression models for estimating EE, we apply unsupervised clustering to implicitly group activities with homogeneous features and generate specific regression models for each activity cluster without requiring manual annotation. The framework therefore does not require specific activity classification, hence eliminating activity type labels. With leave-one-subject-out cross-validation across 10 subjects, an RMSE of 0.96 kcal/min was achieved, which is comparable to the activity-specific model and improves upon a single regression model.

Patient motion training requires adaptive, personalized exercise models and systems that are easy to handle. In this paper, we evaluate a training system based on a smartphone that integrates in clinical routines and serves as a tool for therapist and patient. Only the smartphone's build-in inertial sensors were used to monitor exercise execution and providing acoustic feedback on exercise performance and exercise errors. We used a sinusoidal motion model to exploit the typical repetitive structure of motion exercises. A Teach-mode was used to personalize the system by training under the guidance of a therapist and deriving exercise model parameters. Subsequently, in a Train-mode, the system provides exercise feedback. We validate our approach in a validation with healthy volunteers and in an intervention study with COPD patients. System performance, trainee performance, and feedback efficacy were analysed. We further compare the therapist and training system performances and demonstrate that our approach is viable.

In this work, we introduce methods for studying psychological arousal in naturalistic daily living. We present an activity-aware arousal phase modeling approach that incorporates the additional heart rate (AHR) algorithm to estimate arousal onsets (activations) in the presence of physical activity (PA). In particular, our method filters spurious PA-induced activations from AHR activations, e.g., caused by changes in body posture, using activity primitive patterns and their distributions. Furthermore, our approach includes algorithms for estimating arousal duration and intensity, which are key to arousal assessment. We analyzed the modeling procedure in a participant study with 180 h of unconstrained daily life recordings using a multimodal wearable system comprising two acceleration sensors, a heart rate monitor, and a belt computer. We show how participants' sensor-based arousal phase estimations can be evaluated in relation to daily activity and self-report information. For example, participant-specific arousal was frequently estimated during conversations and yielded highest intensities during office work. We believe that our activity-aware arousal modeling can be used to investigate personal arousal characteristics and introduce novel options for studying human behavior in daily living.

This work introduces an open benchmark dataset to investigate inertial sensor displacement effects in activity recognition. While sensor position displacements such as rotations and translations have been recognised as a key limitation for the deployment of wearable systems, a realistic dataset is lacking. We introduce a concept of gradual sensor displacement conditions, including ideal, self-placement of a user, and mutual displacement deployments. These conditions were analysed in the dataset considering 33 fitness activities, recorded using 9 inertial sensor units from 17 participants. Our statistical analysis of acceleration features quantified relative effects of the displacement conditions. We expect that the dataset can be used to benchmark and compare recognition algorithms in the future.

The development of objective speech quality measures generally involves fitting a model to subjective rating data. A typical data set comprises ratings generated by listening tests performed in different languages and across different laboratories. These factors as well as others, such as the sex and age of the talker, influence the subjective ratings and result in data heterogeneity. We use a linear hierarchical Bayes (HB) structure to account for heterogeneity. To make the structure effective, we develop a variational Bayesian inference for the linear HB structure that approximates not only the posterior over the model parameters, but also the model evidence. Using the approximate model evidence we are able to study and exploit the heterogeneity inducing factors in the Bayesian framework. The new approach yields a simple linear predictor with state-of-the-art predictive performance. Our experiments show that the new method compares favorably with systems based on more complex predictor structures such as ITU-T recommendation P.563, Bayesian MARS, and Gaussian processes.

Due to advancements in lighting technologies new opportunities for application emerge. In this paper an interdisciplinary study towards the development of adaptive lighting environments is presented. An implementation of an adaptive lighting environment is made in the domain of office work. For evaluation, experts from the domains of human-system interaction, activity and context recognition, and system architecture design are interviewed. Contributions are made with regard to the implementation of the adaptive lighting in office environments and an evaluation method that is in line with the interdisciplinary approach. From the evaluation method insights in the fields of human-computer interaction, activity and context recognition and system architecture and (wireless) networking and in topics that span across these fields are gained.

We present an unsupervised speaker identification system for personal annotations of conversations and meetings. The system dynamically learns new speakers and recognizes already known speakers using one audio channel and speech-independent modeling. Multiple personal systems could collaborate in robust unsupervised speaker identification and online learning. The system was optimized for real-time operation on a DSP system that can be worn during daily activities. The system was evaluated on the freely available 24-speaker Augmented Multiparty Interaction dataset. For 5 s recognition time, the system achieves 81% recognition rate. Collaboration between four identification systems resulted in a performance increase of up to 17%, however even two collaborating systems yield an performance improvement. A prototypical wearable DSP implementation could continuously operate for more than 8 hours from a 4.1 Ah battery.

A fundamental challenge limiting information quality obtained from smart sensing garments is the influence of textile movement relative to limbs. We present and validate a comprehensive modeling and simulation framework to predict recognition performance in casual loose-fitting garments. A statistical posture and wrinkle-modeling approach is introduced to simulate sensor orientation errors pertained to local garment wrinkles. A metric was derived to assess fitting, the body-garment mobility. We validated our approach by analyzing simulations of shoulder and elbow rehabilitation postures with respect to experimental data using actual casual garments. Results confirmed congruent performance trends with estimation errors below 4% for all study participants. Our approach allows to estimate the impact of fitting before implementing a garment and performing evaluation studies with it. These simulations revealed critical design parameters for garment prototyping, related to performed body posture, utilized sensing modalities, and garment fitting. We concluded that our modeling approach can substantially expedite design and development of smart garments through early-stage performance analysis.

This paper introduces a sensing garment to support posture coaching in children. The system measures back bending postures using acceleration sensors embedded in the garment. We present a sensing garment architecture and the evaluation of garments of different sizes in a study with 21 children. A vision-based reference system was used to evaluate sensor positions and measurement accuracy for 54 back bending postures and related head positions. Then, we asked eight physiotherapists to rate the children's back postures in this study. Ratings of experts correlated significantly with the back bending measurements obtained from the garment. The garment enables an objective assessment of back postures and could form the basis of a system that provides coaching feedback to improve postural control in children.

Chewing is an essential part of food intake. The analysis and detection of food patterns is an important component of an automatic dietary monitoring system. However chewing is a time-variable process depending on food properties. We present an automated methodology to extract sub-sequences of similar chews from chewing sound recordings. The approach is based on a chew-accurate segmentation of the sound signal, a multi-objective evolutionary search for temporal partitions in the sequence using NSGA-II and a validation of the best solution by classification. We evaluate the method on chewing sound recordings from a four participant study, eating foods with different rheological properties. The proposed methodology allows to determine the most appropriate partitioning of the sequences and extract relevant sound features at the same time. Potato chips and chocolate showed a two-phase structure, for lasagne and apples a single-phase structure was derived. The results led to the hypothesis that a sequential structure can be found in chewing sounds from brittle or rigid foods.