A machine learning based online human motion recognition system with multiple-classifier for exoskeleton

• Published in: IEEE sensors journal Vol. 23; no. 24; p. 1
• Main Authors: Yan, Lingyun, Xiu, Haohua, Wei, Yuyang
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.12.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Activities of daily living; activity detection; Activity recognition; Algorithms; ANN; Artificial neural networks; Exoskeletons; Gait; gait detection; Gait recognition; Human motion; Locomotion; Machine learning; Motion perception; Motion recognition system; Multilayer perceptrons; Neural networks; speed recognition; Walking
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2023.3327723

Motion recognition and classification are crucial for exoskeleton applications in rehabilitation, activities of daily living (ADL), and entertainment. Accurate activity analysis is essential to improve human-machine coupling. However, conventional single-task detection systems, which focus on specific requirements like finite gait events, mode transitions (such as standing-to-sitting), or locomotion speed are inadequate and cannot handle the complex and varied walking environments encountered during ADL. This paper proposes a real-time, multi-classifier system that incorporates three artificial neural network (ANN) models to simultaneously recognize five gait events, nine activities, and walking speeds ranging from 0-8 km/h. Three machine-learning algorithms were fused and utilized to minimize reliance on manual thresholding methods. The activity detection, speed recognition, and gait detection were performed using a 1-Dimension Convolutional Neural Network (1D-CNN), Regression Artificial Neural Network (RANN), and Multilayer Perceptron (MLP), respectively. The experiment was conducted with five subjects wearing a developing cable-driven exoskeleton. The results demonstrate that the proposed portable motion recognition system accurately detected various movements, including gait events with 99.6% accuracy and a time error of 33 ms, recognize speed with a mean square error of 0.12, and activity detection with 96.8% accuracy.