By turning so-called “motion artefacts” into valuable data, researchers reveal that the natural dynamics of loose garments may hold more predictive power than tightly strapped sensors.
Study: Human motion recognition and prediction using loose cloth. Image credit: metamorworks/Shutterstock.com
Most studies require sensors to be firmly attached to the body, but this approach has several disadvantages. A recent paper published in Nature Communications explores the use of sensor-embedded, loose-fitting clothing for human motion analysis.
The rise of smart textiles in movement tracking
The wearable sensor market is projected to reach USD 3.7 billion by 2030, with a compound annual growth rate of 18.3 %. These devices are used across a wide range of sectors, including the diagnosis, monitoring, and treatment of movement-impairing conditions such as Parkinson’s disease and stroke, the tracking of fitness and wellbeing, immersive entertainment experiences in video games and e-sports, and manufacturing applications aimed at reducing workplace accidents, improving productivity, and enabling new forms of human–machine interaction.
Current motion capture technologies rely on wearable sensors, optical systems, or computer-based systems. However, these approaches have notable limitations. Firmly attached sensors can be uncomfortable and inconvenient to remove; they require precise positioning of sensors or markers, and preparation time can be prolonged, especially when multiple sensors are involved. In addition, computer-based motion capture systems often depend on tightly controlled environmental conditions.
E-textiles are fabrics with embedded sensors designed for specific functions. Garments made from such textiles offer a more comfortable and less conspicuous experience than rigidly attached sensors. The present study uses this property to design an improved motion capture system intended for long-term, general use.
Despite this promise, relatively little is known about the movement artefacts that arise when loose clothing shifts independently of the body. Historically, researchers have attempted to minimise such artefacts by tightly bonding sensors to rigid body parts, embedding sensors in tight-fitting garments, or applying machine learning and signal-processing techniques to filter noise from the data.
A smaller body of work has explored the opposite idea: that fabric movement itself might contain useful information about human activity. In this study, the authors provide one of the first systematic investigations into whether loose-cloth motion can be exploited for both motion recognition and motion prediction, aiming to determine whether human movement patterns can be inferred directly from the dynamics of loose garments.
Comparing rigid and garment-mounted sensors
Motion recognition and motion prediction are distinct processes. Motion recognition is the process of identifying the type of movement, such as walking or running, based on past motion data. Motion prediction, by contrast, refers to forecasting the body’s future positions over time.
In this study, movement is classified from a short window of recent fabric-sensor readings using likelihood-based left-to-right hidden Markov models (LR-HMMs). The recognised movement class is then used to predict future motion trajectories.
The authors hypothesised that clothing movement could enhance recognition accuracy compared with rigid body measurements, particularly when distinguishing between very similar movement classes. Importantly, they proposed that this improvement could be achieved using a much shorter history of past motion data.
This hypothesis rests on the idea that fabric behaves as a flexible, nonlinear system rather than as a rigid extension of the body. Because clothing can move in multiple directions and exhibit complex dynamics, it may amplify subtle differences between movement classes, thereby increasing the information available for classification and prediction.
To test this, the researchers used LR-HMMs to model motion based on orientation sensor data. Although the primary results focus on this modelling approach, the findings were also replicated in supplementary analyses using other modelling strategies and sensing modalities.
LR-HMMs were selected because they are straightforward to construct and train, and they use optimised algorithms for scoring time-series data. Orientation sensing was chosen because it is the primary modality used in accelerometers and inertial measurement units (IMUs), which are commonly found in wearable and untethered motion capture systems.
The team compared tightly attached sensors with loosely attached, fabric-mounted sensors across tasks of varying complexity, frequency, pattern, and velocity. Unidimensional movements were generated using a scotch yoke mechanism, while discrete multidimensional movements were produced with a robot arm. Human reaching movements were also recorded from 22 participants, with 50 samples collected for each of four target buttons, starting from a central position.
Fabric dynamics boost accuracy in hard tasks
Across experimental tasks, fabric movements predicted unidimensional body movements at varying frequencies with greater accuracy and efficiency than rigidly attached sensors. Similar advantages were observed for irregular, multidimensional movements.
Fabric-mounted sensors increased motion recognition accuracy by up to 40 % and reduced the required history of past movement by approximately 80 % compared with body-attached sensors. The benefits were most pronounced when distinguishing between highly similar movement classes and became more evident at higher movement speeds.
To quantify this improvement, the researchers measured cross-fitness distance, a statistical measure of separability between movement models, over time. Greater cross-fitness distance corresponds to higher information content and improved differentiation between movement classes. Fabric signals consistently produced larger separations than rigid-body signals, with faster rigid-body motion generating greater information content in the fabric response.
Human reaching movements were also predicted with greater accuracy when using loose clothing fitted with sensors, although these results were obtained under controlled laboratory conditions rather than in free-living environments.
For more difficult prediction tasks, fabric-mounted sensors performed best and did so largely independently of precise sensor positioning, though the underlying mechanisms remain unclear. In simpler tasks, combining fabric-mounted and rigidly attached sensors produced the highest accuracy.
The findings reinforce the practical advantages of e-textiles, suggesting that they may improve user acceptance while expanding the potential for accurate motion capture using ordinary clothing.
Environmental testing showed that airflow velocity influenced fabric behaviour and reduced recognition accuracy. Low airflow caused only slight degradation, primarily in complex tasks, whereas higher airflow led to more pronounced deterioration compared with rigid sensors. In contrast, variations in room temperature and humidity did not significantly affect performance, suggesting that these factors may be manageable in practical applications.
Overall, the results challenge the conventional assumption that tightly attached sensors are inherently superior for motion prediction, at least under the controlled experimental conditions tested.
Limitations and future directions
One notable limitation is that most experiments were conducted using a tethered system, whereas most real-world motion capture applications require portable, wearable systems. Although the authors conducted a preliminary validation using IMUs to demonstrate proof of concept, this did not constitute a full independent field validation.
The study also relied on a relatively simple modelling framework, meaning that further research is needed to confirm and extend the findings using more complex models and in real-world settings.
Future work could investigate why sensor position and distance from the rigid attachment had a limited influence on performance, potentially exploring harmonic or resonance-like effects in fabric dynamics. Such research may help define an optimal degree of looseness for different applications.
The authors also draw conceptual parallels to biological and cultural phenomena, noting that animals and humans often use flexible appendages or flowing materials, such as spider webs or loose garments, that encode rich environmental information. However, these analogies remain conceptual and were not directly tested in the present study.
Rethinking tight attachment in wearable design
The study suggests that garment motion can serve as a valuable and previously underappreciated source of information for analysing human movement. By harnessing the nonlinear dynamics of loose fabrics, researchers may be able to design more comfortable and efficient intelligent textiles.
Further validation in portable, real-world systems will be necessary before such approaches can be widely adopted.
Download your PDF copy by clicking here!
Dual targeting strategy suppresses pancreatic cancer cell growth