Human Movement Forecasting with Loose Clothing

TL;DR

This work compares motion forecasting from fabric-attached versus rigid body-attached sensors, showing that loose clothing can yield superior predictive performance. It formulates an lr-HMM framework with Baum–Welch training and Viterbi decoding to classify motion class from fabric data and forecast the body’s future trajectory, validated through robotic-arm and human-reaching experiments. The key findings are that fabric-attached sensors can improve prediction accuracy by up to about 40% and reduce the required history length by about 80% to reach high accuracy, with higher cross-fitness distances indicating more discriminative information in fabric motion. The results imply broad implications for wearable sensing and human–robot collaboration, suggesting that deliberately leveraging clothing motion artefacts can enhance predictive capabilities in real-world settings.

Abstract

Human motion prediction and trajectory forecasting are essential in human motion analysis. Nowadays, sensors can be seamlessly integrated into clothing using cutting-edge electronic textile (e-textile) technology, allowing long-term recording of human movements outside the laboratory. Motivated by the recent findings that clothing-attached sensors can achieve higher activity recognition accuracy than body-attached sensors. This work investigates the performance of human motion prediction using clothing-attached sensors compared with body-attached sensors. It reports experiments in which statistical models learnt from the movement of loose clothing are used to predict motion patterns of the body of robotically simulated and real human behaviours. Counterintuitively, the results show that fabric-attached sensors can have better motion prediction performance than rigid-attached sensors. Specifically, The fabric-attached sensor can improve the accuracy up to 40% and requires up to 80% less duration of the past trajectory to achieve high prediction accuracy (i.e., 95%) compared to the rigid-attached sensor.

Figures (5)

• Figure 1: Illustration of human movement forecasting. The future human movement (dashed line and shaded area) is forecasted based on the past movement of the clothing-attached sensor (solid line).
• Figure 2: The framework for motion prediction and trajectory forecasting based on . The future human movement $\overset{\rightarrow}{\mathbf{Y}^r}$ (dashed line and shaded area) is forecasted based on the prediction label $c$ of the past clothing movement $\overset{\leftarrow}{\mathbf{Y}^f}$ (solid line) and the probability trajectory model $\tilde{\boldsymbol{\theta}}_{c}^r$ formulated by body movement using the Baum-Welch algorithm and the Viterbi algorithm.
• Figure 3: Experimental setup. The robot arm with a piece of fabric attached to the end effector. The actual moving trajectories and the longitudinal/X-axis (roll) orientation of the rigid-attached sensor and the fabric-attached sensor duringlinear andcurved movement. Note that the variations of rigid movements in both types of movements are less obvious.
• Figure 4: The accuracy of motion prediction and the cross-fitness distance from the first time step to $1.2 s$ in the contexts of linear andcurved movements. Reported are the mean value $\pm$$\text{s.d}/5$. Panelsandshow the forecasting of linear and curved movements, respectively. The future robot movement is forecasted (dashed line) based on $1s$ of past movement (solid line). The future trajectory shown is the mean value (dashed line) $\pm$$\text{s.d}\times5$ (shaded area).
• Figure 5: Experimental setup. The participant presses a start button and then reaches for one of the target buttons. The sensors are attached to the wrist and the tip of the sleeve. The motion prediction accuracy and the cross-fitness distance betweenbutton $1$ and $2$,button $2$ and $3$ andbutton $3$ and $4$ given various durations of the past movement. Reported are the mean value $\pm$$\text{s.d}/5$.The vertical position of the sensor's movement when the human arm is reaching button $2$ and button $3$.The future movement of the human body (dashed line and shaded area) is predicted based on their past movement (solid line) from the initial time step up to 0.6s. The future trajectory shown is the mean value (dashed line) $\pm$$\text{s.d}\times5$ (shaded area). The motion prediction accuracies using the past movement until $1s$ recorded by sensor $R_{1}$ and $F_{4}$ are shown.