Although, gaze-based interaction has been investigated since the 1980s and provides promising concepts to realize cognitive systems and support universal interaction within distributed environments, the main challenges, such as the Midas touch problem [16] or calibration are still frequent topics of research. In this work, Natural Pursuit Calibration is presented, which is a comfortable, unobtrusive technique enabling ongoing attention detection and eye tracker calibration within an off-screen context. The user is able to perform calibration, without a digital user interface, artificial annotation of the environment nor further assistance, by simply following any arbitrary moving target. Due to the characteristics of the calibration process, it can be executed simultaneously to any primary task, without active user participation. A two-stage evaluation process is conducted to (i) optimize parameter settings in a first setup and (ii) compare the accuracy as well as the user acceptance of the proposed procedure to prevailing calibration techniques.

Depth cameras have been known to be capable of picking up the small changes in distance from users’ torsos, to estimate respiration rate. Several studies have shown that under cer- tain conditions, the respiration rate from a non-mobile user facing the camera can be accurately estimated from parts of the depth data. It is however to date not clear, what factors might hinder the application of this technology in any setting, what areas of the torso need to be observed, and how readings are affected for persons at larger distances from the RGB-D camera. In this paper, we present a benchmark dataset that consists of the point cloud data from a depth camera, which monitors 7 volunteers at variable distances, for variable methods to pin-point the person’s torso, and at variable breathing rates. Our findings show that the respiration signal’s signal-to-noise ratio becomes debilitating as the distance to the person approaches 4 metres, and that bigger windows over the person’s chest work particularly well. The sampling rate of the depth camera was also found to impact the signal’s quality significantly.

Photoplethysmography (PPG) sensors have become a prevalent feature included in current wearables, as the cost and size of current PPG modules have dropped significantly. Research in the analysis of PPG data has recently expanded beyond the fast and accurate characterization of heart rate, into the adaptive handling of artifacts within the signal and even the capturing of respiration rate. In this paper, we instead explore using state-of-the-art PPG sensor modules for long-term wearable deployment and the observation of trends over minutes, rather than seconds. By focusing specifically on lowering the sampling rate and via analysis of the spectrum of frequencies alone, our approach minimizes the costly illumination-based sensing and can be used to detect the dominant frequencies of heart rate and respiration rate, but also enables to infer on activity of the sympathetic nervous system. We show in two experiments that such detections and measurements can still be achieved at low sampling rates down to 10 Hz, within a power-efficient platform. This approach enables miniature sensor designs that monitor average heart rate, respiration rate, and sympathetic nerve activity over longer stretches of time.

We present a novel Smartwatch-based approach, to enable Mobile Assisted Living (MAL) for users with special needs. A major focus group for this approach are elderly people. We developed a tool for caretakers applicable in home environments, nursing care, and hospitals, to assess the vitality of their patients. Hereby, we particularly focus on the prediction of falls, because falls are a major reason for serious injuries and premature death among elderly. Therefore, we propose a multi parametric score based on standardized fall risk assessment tests, as well as on sleep quality, medication, patient history, motor skills, and environmental factors. The resulting total fall risk score reflects individual changes in behavior and vitality, which consequently enables for fall preventing interventions. Our system has been deployed and evaluated in a pilot study among 30 elderly patients over a period of four weeks.

The use of publicly available repositories containing movement traces of real or experimental subjects is a key aspect to define an evaluation framework that allows a systematic assessment of wearable fall detection systems. This papers presents a detailed analysis of a public dataset of traces which employed five sensing points to characterize the user’s mobility during the execution of ADLs (Activities of Daily Living) and emulated falls. The analysis is aimed at analysing two main factors: the importance of the election of the position of the sensor and the possible impact of the user’s personal features on the statistical characterization of the movements. Results reveal the importance of the nature of the ADL for the effectiveness of the discrimination of the falls.

Step detection with accelerometers is a very common feature smart wearables already include. However, when using a wheeled walking frame / rollator, current algorithms may be of limited use, since a different type of motion is being excreted. In this paper we uncover these limitations of current wearables by a pilot study. Furthermore, we investigated an accelerometer-based step detection when using a wheeled walking frame, when mounting an accelerometer to the frame and at the user’s wrist. Our findings include knowledge on the spreading of signal on different axis, knowledge on the required sensor quality and knowledge on the impact of different floor types. In conclusion, we outline a new step detection algorithm based on accelerometer input data. Our algorithm can significantly empower future off-the-shelf wearables with the capability to sufficiently detect steps with elderly people using a wheeled walking frame. This can help to evaluate the state of health with regard to the human motor system and even to determine the progress of certain diseases.

Inertial Measurement Units (IMUs) have proven to be a promising candidate for joint kinematics assessment during human locomotion. The benefits associated with IMU-based joint angle measurements are ease of handling, flexibility and low cost. However, a known limitation is that the joint axes in terms of the attached IMUs need to be identified in order to decompose IMU measurements into joint angles. Conventionally, careful alignment of the IMUs with respect to the body segments and/or calibration motions are required. In this paper, a novel approach is proposed to estimate the joint axes of the hip and knee joint during gait. Our method is easy to use, self-calibrating and real-time capable using the obtained IMU data during gait. In addition to prior methods, the algorithm profits from the periodicity during gait in order to deal with three (rotational) degrees of freedom (3-DoF) motions. Experiments with 8 healthy subjects walking on a motor-driven treadmill have been conducted. The joint axes converged onto the expected axes in all trials and the convergence times averaged less than 15 seconds.

Semantic descriptions or attribute representations have been used successfully for object and scene recognition, and for word-spotting. However, these representations have not been explored deeply on human activity recognition (HAR). Particularly, in the manual order picking process, attribute representations are beneficial for dealing with the versatility of activities in the process. This paper compares the performance of deep architectures trained using different attribute representations for HAR. Besides, it evaluates their quality from the perspective of practical application.

This paper presents preliminary results of the ongoing project TheOdor which explores the potential of electronic noses that make use of commodity gas sensors (MOS, MEMS) for applications in the smarthome, for example to classify human activities based on the odors that activities generate. We describe the system and its components and report on classification results from first validation experiments.

Motion Capture (MOCAP) Systems have been used to analyse body motion and postures in biomedicine, sports, rehabilitation, and music. With the aim to compare the precision of low-cost devices for motion tracking (e.g. \textit{Myo}) with the precision of MOCAP systems in the context of music performance, we recorded MOCAP and Myo data of a top professional violinist executing four fundental bowing technques (i.e. \textit{Détaché}, \textit{Martelé}, \textit{Spiccato} and \textit{Ricochet}). Using the recorded data we applied machine learning techniques to train models to classify the four bowing techniques. Despite intrinsic differences between the MOCAP and low-cost data, the Myo-based clasifier resulted in slightly higher accuracy than the MOCAP-based classifier. This result shows that it is possible to develop music-gesture learning applications based on low-cost technology which can be used in home environments for self-learning practitioners.

Game play in the sport of basketball tends to combine highly dynamic phases in which the teams strategically move across the field, with specific actions made by individual players. Analysis of basketball games usually focuses on the locations of players at particular points in the game, whereas the capture of what actions the players were performing remains underrepresented. In this paper, we present an approach that allows to monitor players’ actions during a game (such as dribbling, shooting, blocking, or passing) with wrist-worn inertial sensors. In a our study, inertial data from a sensor worn on the wrist were recorded during training and game sessions from three players. We illustrate that common features and classifiers are able to recognize short actions, with overall accuracy performances around 83.6% (k-Nearest- Neighbor) and 87.5% (Random Forest). Some actions, such as jump shots performed well (±95% accuracy), whereas some types of dribbling achieving low (±44%) recall.

Human activity recognition using inertial sensors is a feature many people use in their smartphones or smartwatches, providing information on sports and physical activities of each individual. But while the position a smartphone is worn in varies between persons and circumstances, a smartwatch moves constantly, in rhythm with its user's arms. Both problems can make an activity recognition less reliable. Attaching an inertial sensor to the head provides reliable information on the movements of the whole body while not being superimposed by many additional movements. This can be achieved by fixing sensors to glasses, helmets, or headphones. In this paper, we present a system using head-mounted inertial sensors for human activity recognition. We compare it to existing research work and show possible advantages or disadvantages of positioning a single sensor on the head to recognize physical activities.

Human Activity Recognition (HAR) plays an important role in many real world applications. Currently, various techniques have been proposed for sensor-based "HAR" in daily health monitoring, rehabilitative training and disease prevention. However, non-visual sensors in general and wearable sensors in specific have several limitations: acceptability and willingness to use wearable sensors; battery life; ease of use; size and effectiveness of the sensors. Therefore, adopting vision-based human activity recognition approach is more viable option since its diversity would enable the application to be deployed in wide range of domains. The most popular technique of vision based activity recognition, Deep Learning, however, requires huge domain-specific datasets for training which, is time consuming and expensive. To address this problem this paper proposes a Transfer Learning technique by adopting vision-based approach to "HAR" by using already trained Deep Learning models. A new stochastic model is developed by borrowing the concept of "Dirichlet Alloaction" from Latent Dirichlet Allocation (LDA) for an inference of the posterior distribution of the variables relating the deep learning classifiers predicted labels with the corresponding activities. Results show that an average accuracy of 95.43% is achieved during training the model as compared to 74.88 and 61.4% of Decision Tree and SVM respectively. However, testing accuracy suffers due to non-reproducibility of labels by the deep learning models.

In this paper, a new activity recognition technique is introduced based on the gray level co-occurrence matrices (GLCM) from a 3D dense optical flow of the input RGB and Depth videos. These matrices are one of the earliest techniques used for image texture analysis which are representing the distribution of the intensities and information about relative positions of neighboring pixels of an image. In this work, we propose a new method to extract feature vector values using the well-known Haralick features from GLCM matrices to describe the flow pattern by measuring meaningful properties such as energy, contrast, homogeneity, entropy, means, correlation and sum average to capture local spatial and temporal characteristics of the motion through the neighboring optical flow orientation and magnitude. To evaluate the proposed method and improve the activity recognition problem, we apply a recognition pipeline that involves the bag of local spatial and temporal features and three types of machine learning classifiers are used for comparing the recognition accuracy rate of our method. These classifiers are random forest, support vector machine and K-nearest neighbor. The experimental results carried on two well-known datasets (Gaming datasets (G3D) and Cornell Activity Datasets (CAD-60)), which demonstrate that our method outperforms the results achieved by several widely employed spatial and temporal feature descriptors methods.

Human activity recognition using wearable sensors is an area of interest for various domains like healthcare, surveillance etc. Various approaches have been used to solve the problem of activity recognition. Recently deep learning methods like RNNs and LSTMs have been used for this task. But these architectures are unable to capture long term dependencies in time series data. In this work, we propose to use the Temporal Convolutional Network architecture for recognizing the activities from the sensor data obtained from a smartphone. Due to the potential of the architecture to take variable length input sequences along with significantly better ability to capture long term dependencies, it performs better than other deep learning methods. The results of the proposed methods shows an improved performance over the existing methods.