Lower Limb Locomotion Modes Recognition Based on Multiple-source Information and General Regression Neural Network

In order to improve recognition rate of lower limb locomotion modes, a method based on multiple-source information and general regression neural network (GRNN) is proposed. Users' daily lower limb locomotion modes are decomposed into different segments to form the recognition goals using the plantar pressure sensor. For surface electromyography (sEMG) signal, three features are used, i.e. skewness, kurtosis, and power spectral entropy. The hip joint angle is chosen as leg features. In order to decrease the time for training the models and to prevent overfitting, principal component analysis (PCA) is used to reduce the dimension of the extracted features. GRNN is used to recognize 3 kinds of motions, namely stairs ascent, stairs descent and level-ground walking. The experimental results show that the recognition correct rate is 90.16%.