AI Summary & Key Takeaways
AI InsightThese results show that a suitably designed CNN can directly regress kinematic variables from raw inertial data, and that input channel organisation strongly influences performance, underscoring the importance of input structuring even in end-to-end deep learning, particularly for biomechanical time series with strongly co-varying signals.
📌 3 Key Takeaways
Abstract
Accurate gait analysis is fundamental to assessing mobility, diagnosing neurological disorders and monitoring rehabilitation progress. Optical motion capture remains the gold standard for kinematic parameters, but its high cost, need for specialised laboratories and limited capture volume restrict clinical use. Wearable inertial measurement units offer a low-cost alternative for continuous monitoring in natural environments, yet deriving clinically meaningful kinematic variables from raw accelerometer and gyroscope signals remains a signal processing challenge. Deep learning, particularly convolutional neural networks (CNNs), can learn hierarchical features directly from raw sensor data, potentially eliminating hand-crafted feature engineering. This study investigates the impact of input channel grouping on the prediction performance of a bespoke CNN architecture for kinematic signal regression from inertial data, specifically for predicting the stride length normalised to body height. We used data from the Machine Learning and Data Analytics Laboratory at Friedrich-Alexander-Universität Erlangen-Nürnberg, comprising 101 gait trials with synchronised triaxial accelerometer and gyroscope data from both ankles and reference stride length from a GAITRite walkway. The CNN took a 400×12 input matrix: 12 sensor signals (6 per ankle), 400 time steps. Two input representations were compared: (i) standard stacking in the original channel order, and (ii) a novel grouped representation with all X-axis (primary gait direction) signals placed contiguously to facilitate cross-channel feature learning by the convolutional filters. Models were trained on 91 trials and tested on 10, using root-mean-square error (RMSE) as primary metric, comparing mini-batch size 30 against full-batch. The grouped representation achieved a substantially lower RMSE (0.17) than the standard one (0.21), a 19% relative improvement. Mini-batch size affected learning-curve smoothness but not final RMSE. These results show that a suitably designed CNN can directly regress kinematic variables from raw inertial data, and that input channel organisation strongly influences performance, underscoring the importance of input structuring even in end-to-end deep learning, particularly for biomechanical time series with strongly co-varying signals. Task-driven channel reorganisation offers practical guidance for deep learning in wearable sensor-based movement analysis.
Özet
Accurate gait analysis is fundamental to assessing mobility, diagnosing neurological disorders and monitoring rehabilitation progress. Optical motion capture remains the gold standard for kinematic parameters, but its high cost, need for specialised laboratories and limited capture volume restrict clinical use. Wearable inertial measurement units offer a low-cost alternative for continuous monitoring in natural environments, yet deriving clinically meaningful kinematic variables from raw accelerometer and gyroscope signals remains a signal processing challenge. Deep learning, particularly convolutional neural networks (CNNs), can learn hierarchical features directly from raw sensor data, potentially eliminating hand-crafted feature engineering. This study investigates the impact of input channel grouping on the prediction performance of a bespoke CNN architecture for kinematic signal regression from inertial data, specifically for predicting the stride length normalised to body height. We used data from the Machine Learning and Data Analytics Laboratory at Friedrich-Alexander-Universität Erlangen-Nürnberg, comprising 101 gait trials with synchronised triaxial accelerometer and gyroscope data from both ankles and reference stride length from a GAITRite walkway. The CNN took a 400×12 input matrix: 12 sensor signals (6 per ankle), 400 time steps. Two input representations were compared: (i) standard stacking in the original channel order, and (ii) a novel grouped representation with all X-axis (primary gait direction) signals placed contiguously to facilitate cross-channel feature learning by the convolutional filters. Models were trained on 91 trials and tested on 10, using root-mean-square error (RMSE) as primary metric, comparing mini-batch size 30 against full-batch. The grouped representation achieved a substantially lower RMSE (0.17) than the standard one (0.21), a 19% relative improvement. Mini-batch size affected learning-curve smoothness but not final RMSE. These results show that a suitably designed CNN can directly regress kinematic variables from raw inertial data, and that input channel organisation strongly influences performance, underscoring the importance of input structuring even in end-to-end deep learning, particularly for biomechanical time series with strongly co-varying signals. Task-driven channel reorganisation offers practical guidance for deep learning in wearable sensor-based movement analysis.