Human Locomotion Classification for Different Terrains Using Machine Learning Techniques

Abstract

Gait analysis on healthy subjects was performed based on surface electromyographic and acceleration sensor signal, implemented through machine learning approaches. The surface EMG and 3-axes acceleration signals have been acquired for 5 different terrains: level ground, ramp ascent, ramp descent, stair ascent, and stair descent. These signals were acquired from the tibialis anterior and gastrocnemius medial head muscles that correspond to dorsiflexion and plantar flexion, respectively. After feature extraction, these signals are fed to 5 conventional classifiers: linear discriminant analysis, k-nearest neighbors, decision tree, random forest, and support vector machine, that classify different terrains for human locomotion. We compared the classification results for the above classifiers with deep neural network classifier. The objective was to obtain the features and classifiers that are able to discriminate between 5 locomotion terrains with maximum classification accuracy in minimum time by acquiring the signal from the least number of leg muscles. The results indicated that the support vector machine gives the highest classification accuracy of 99.20 (± 0.80)% for the dataset acquired from 15 healthy subjects. In terms of both accuracy and computation time, the support vector machine outperforms other classifiers.