Recognition of Lower Limb Movements Using Machine Learning Methods and Bispectral Maps of Wireless sEMG Measurements

In this letter, an attempt is made to develop a movement recognition framework for the classification of five locomotion activities using bispectral representation of wireless surface electromyography (sEMG) measurements and machine learning classifiers. For this purpose, wireless sEMG signals are recorded during locomotive activities, namely, level ground walking, ramp up, ramp down, stair up, and stair down from six muscles of the lower limbs, namely gastrocnemius, semitendinosus, rectus femoris, biceps femoris, tibialis anterior, and vastus lateralis. Nine features, namely variance, root mean square, integrated EMG, mean absolute value, waveform length, Willison amplitude, skewness, entropy, and zero-crossing, are extracted from the recorded sEMG. Tree-based machine learning approaches namely random forest and extreme gradient boosting (XG Boost) are employed to classify lower limb movements. Moreover, we explore the synergy of bispectral maps from sEMG data with deep learning models to enhance the discrimination of lower limb activities. The proposed approach effectively discriminates various lower limb movements. The evaluation of machine learning models demonstrates promising results, with random forest and XG Boost achieving notable accuracy rates. Notably, the convolutional neural network model outperforms others with a testing accuracy of 91.33%, closely followed by the visual geometry group with 16 layers model at 89.9%. The results indicate the potential utility of the proposed approach in leg-machine interface development and rehabilitation for lower limb amputees. This research contributes to the advancement of movement recognition technology, offering valuable insights for biomedical engineering and rehabilitation sciences.