Holger Harms

Recognizing Upper Body Postures using Textile Strain Sensors

In this paper we present a garment prototype using strain sensors to recognize upper body postures. A novel thermoplastic elastomer strain sensor was used for measuring strain in the clothing. This sensor has a linear resistance response to strain, a small hysteresis and can be fully integrated into textile. A study was conducted with eight participants wearing the garment and performing a total of 27 upper body postures. A Naive Bayes classification was applied to identify the different postures. Nearly a complete recognition rate of 97% was achieved when the classification was adapted to the individual participant. A classification rate of 84% was achieved for an all-user classification and 65% for an independent user. These results show the feasibility to recognize postures with our setup, even in an unseen user setting. Furthermore, we used the garment prototype in a gym experiment to explore its potential for rehabilitation and fitness training. Intensity, speed and number of repetitions could be obtained from the garment sensor data.

ETHOS: Miniature orientation sensor for wearable human motion analysis

Inertial and magnetic sensors offers a sourceless and mobile option to obtain body posture and motion for personal sports or healthcare assistants, if sensors could be unobtrusively integrated in casual garments and accessories. We present in this paper design, implementation, and evaluation results for a novel miniature attitude and heading reference system (AHRS) named ETHOS using current off-the-shelf technologies. ETHOS has a unit size of 2.5cm3, which is substantially below most currently marketed attitude heading reference systems, while the unit contains processing resources to estimate its orientation online. Results on power consumption in relation to sampling frequency and sensor use are presented. Moreover two sensor fusion algorithms to estimate orientation: a quaternion-based Kalman-, and a complementary filter. Evaluations of orientation estimation accuracy in static and dynamic conditions revealed that complementary filtering reached sufficient accuracy while consuming 46% of a Kalmans power. The system runtime of ETHOS was found to be 10 hours at a complementary filter update rate of 128Hz. Furthermore, we found that a ETHOS prototype functioned with a sufficient accuracy in estimating human movement in real-life conditions using an arm rehabilitation robot.

SMASH: a distributed sensing and processing garment for the classification of upper body postures

This paper introduces a smart textile for posture classification. A distributed sensing and processing architecture is implemented into a loose fitting long sleeve shirt. Standardized interfaces to remote periphery support the variable placement of different sensor modalities at any location of the textile. The shirt is equipped with acceleration sensors in order to determine the postural resolution and the systems feasibility for applications in movement rehabilitation. For the garment characterization an arm posture measurement method is proposed and applied in a study with 5 users. The classification performance is analyzed on data from overall 8 users, conducting 12 posture types, relevant for shoulder and elbow joint rehabilitation. We present results for different user-modes, with classification rates of 89% for a user-independent evaluation. Moreover, the relation of body dimensions on the posture classification performance are analyzed.

Monitoring Kinematic Changes With Fatigue in Running Using Body-Worn Sensors

In this paper, we investigate monitoring of kinematic changes evoked by fatigue in running using wearable technology. Movement data were recorded with ETHOS devices. ETHOS is the ETH Orientation Sensor, a customized inertial measurement unit for unconstrained monitoring of human movement. We perform two real-world experiments, in which 21 runners of different skill levels participated. The real-world experiments capture two exhausting 45-min runs: one on a treadmill and one on a conventional outdoor track. We describe and evaluate algorithms to extract kinematic parameters from the sensor data. We identified parameters that change with fatigue for all runners, ones that change for runners of distinct skill levels, and ones that are dependent on an individuals running technique. Overall, we found that observations from treadmill running are not always generalizable to outdoor running. We, thus, argue for using wearable technology to provide athletes and trainers with continuous, quantitative objective measurements of running technique. These could be used to further gain insight into the complex relationship of running kinematics, injury risk, fatigue, and running economy.

Textile pressure sensors for sports applications

Wearable sports trainers are built upon sensor systems recognizing the activities performed by its users. In snowboarding, one of the fastest growing sports in the world, traditional activity recognition approaches make use of pressure insoles with force-sensitive resistors, which, however, are particularly uncomfortable to wear. To make these measurements more convenient, we have developed textile pressure sensors using the principle of a variable capacitor. Electrodes of conductive textiles coated with silver arranged on both sides of compressible spacers made from Croslite™ form a capacitor, whose capacitance indicates the applied pressure. We integrated three sensors into a snowboarding sock at relevant positions under the heel and the ball of the foot. Outdoor experiments on a ski slope in the Matterhorn Glacier Paradise (Zermatt, Switzerland) show that the machine learning algorithm NCC can detect turns, the basic activity of snowboarding, from the sensor data with an accuracy of 84 percent. Moreover, indoor experiments reveal that NCC can clearly distinguish whether a person wearing our sensor socks is standing on the ball of the foot, flat or on the heel. These results suggest the socks might also be used for gait analysis or the monitoring of the in-shoe pressure distribution of runners.

Rapid prototyping of smart garments for activity-aware applications

Continuous miniaturization of electronics and sensing elements stimulate the evolution of novel unobtrusively integrated smart garments that sense their environment and provide personalized assistance to its wearer. The development of smart garments requires robust integration techniques for electronics and textiles in one common system. Furthermore, recognition algorithms are needed to derive information on the wearers activity and context within the smart garment. In this work both challenges are addressed in a smart shirt system, called SMASH. SMASH was developed as a rapid prototyping system for smart garment developments. We introduced in this work our approach for prototyping smart garments and present design, implementation, and evaluation of SMASH. The SMASH system embeds a distributed hierarchical architecture of sensing and processing functions in an off-the-shelf long-sleeve shirt. The system design focused on scalability regarding sensors and processing resources, as well as rapid deployment in different applications. We demonstrated the versatility of SMASH in three application evaluations that represent different prototyping phases of smart garments. For these studies several systems of different sizes were implemented. The SMASH system helps to bypass time- and cost-intensive implementation iterations using multiple garment prototypes.

Influence of a loose-fitting sensing garment on posture recognition in rehabilitation

Several smart sensing garments have been proposed for postural and movement rehabilitation. Existing systems require a tight-fitting of the garment at body segments and precise sensor positioning. In this work, we analyzed errors of a loose-fitting sensing garment on the automatic recognition of 21 postures, relevant in shoulder and elbow-rehabilitation. The recognition performance of garment-attached acceleration sensors and additional skin-attached references was compared to discuss challenges in a garment-based classification of postures. The analysis was done with one fixed-size shirt worn by seven participants of varying body proportions. The classification accuracy using data from garment-integrated sensors was on average 13% lower compared to that of skin-attached reference sensors. This relation remained constant even after selecting an optimal input feature set. For garment-attached sensors, we observed that the loss in classification accuracy decreased, if the body dimension increased. Moreover, the alignment error of individual postures was analyzed, to identify movements and postures that are particularly affected by garment fitting aspects. Contrarily, we showed that 14 of the 21 rehabilitation-relevant postures result in a low sensor alignment error. We believe that these results indicate critical design aspects for the deployment of comfortable garments in movement rehabilitation and should be considered in garment and posture selection.

What Do Sensors Know about Your Running Performance

About 489 000 athletes have finished a marathon in the US in 2009, the average training distance of athletes is 47.5 km per week. While average fitness runners (55.8%) train by self perception, a systematic assessment of kinematic parameters is limited to elite athletes that have access to instrumented environments. This work investigates the potential of wearable sensors to derive kinematic features in running. We equipped 12 runners of different performance levels with each 12 miniature ETHOS units. ETHOS constitutes a miniature inertial measurement unit (IMU) that is optimized for long term monitoring in unconstrained environments. We found that a minimum set of two acceleration sensors attached to the athletes foot and hip is sufficient to derive kinematic features that allow for distinction of experienced and unexperienced runners. Our work constitutes a first step towards personal training assistance providing runners kinematic metrics for performance improvement and injury risk reduction.

Estimating Posture-Recognition Performance in Sensing Garments Using Geometric Wrinkle Modeling

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.

Modeling and simulation of sensor orientation errors in garments

We report in this paper on a novel modeling and simulation approach to predict orientation errors of garment-attached sensors and their effect on posture classification. Such errors occur frequently in smart garment implementations and can reduce sensor information quality for movement and posture recognition. A kinematic model of the human upper-body was developed to simulate upper limb postures and the output of virtual 3D acceleration sensors. The model was enhanced with a statistical approximation of garment-related orientation errors. We derived this model from acceleration sensor deviations between skin- and garment-attached units. The feasibility of our body model and the garment-attached sensor deviation was validated in experimental data. We compared the classification performance for ten posture types that are frequently used in shoulder rehabilitation. In a validation set of 7 participants we observed similar classifier confusions and a relative error of 2.6% (SD:±3.2%) between simulation and experiment. We utilized the model to estimate classification performance for further simulated textile error distributions. Our simulations showed that classification performance depends on low deviations of an acceleration sensor at the lower arm, while a sensor at the upper arm was less critical. Moreover, we included magnetic field sensors in our simulation. With the help of this additional modality our posture classification performance increased by 18%. We conclude that simulation of skin- and garment-attached sensors is a feasible approach to expedite design and development process of smart garments.