Recognition of Lower Limb Motions via Surface Electromyography

Wearable exoskeleton can help people with mobility impairments, such as stroke and amputation patients, to improve their rehabilitation. Traditional exoskeleton control signals include plantar pressure and joint angle. These signals can only reflect the current state and human motion, but cannot predict the motion. As electromyography (EMG) signal occurs before the motion, it can be used as the input signal to predict the subject's motion intention. In this paper, EMG signals are used to identify human lower limb motions, and different algorithms are compared to find the most suitable one for motion recognition. Seven types of lower limb motions are involved, i.e., walking, going upstairs, going downstairs, walking uphill, walking downhill, squatting, and standing. The surface EMG signals from 6 muscles are measured, i.e., rectus femoris, vastus lateralis, semitendinosus, biceps femoris, gastrocnemius, and tibial anterior. Butterworth filter and wavelet threshold are used to denoise the raw EMG signals, and support vector machine (SVM), C4.5 decision tree, and backpropagation (BP) neural network are used to recognize the different motions. To improve the accuracy and speed of recognition, principal component analysis (PCA) is adopted to reduce the feature dimensionality. Before dimensionality reduction, the BP neural network shows the highest accuracy, i.e., up to 97.32%, but it takes 17.205 seconds on average. After dimensionality reduction, the computational cost of each algorithm is significantly reduced; and instead, the SVM shows the highest accuracy, i.e., up to 97.44%, and it takes only 0.004 seconds. The results show that the combination of SVM and PCA has the best performance in lower limb motion recognition.