sEMG Sensor-Based Human Lower Limb Activity Recognition Using Machine Learning Algorithms

Abstract

Human lower limb activity recognition focuses on determining the activities of a person by monitoring their actions on the basis of datasets acquired via sensors such as accelerometers, gyroscopes, surface electromyography (sEMG), etc. sEMG is a computer-aided approach that incorporates useful information regarding movements of limbs and is also used for analyzing and recording the electrical activity generated by skeletal muscles. This paper demonstrates the analysis of the sEMG sensor-based dataset obtained from different muscles of 22 subjects performing activities such as walking, sitting, and standing. Out of these subjects, 11 seemed normal and the rest exhibited abnormalities. As a consequence of unprocessed data, discrete wavelet transform is applied to denoise the signal. Further, the overlapping windowing approach is used to execute the signal’s segmentation, followed by the procedure of feature extraction, which is carried out by extracting five-time domain features. Several machine learning models, such as random forest, gradient boosting, k-nearest neighbors, support vector machine using radial basis function, and the polynomial kernel were implemented. The results show that random forest, having cross-validation of 5-fold, achieved the best accuracy for normal (85.68%) and abnormal subjects (83.96%) in determining human activity.