A comparative study on human loco-motor activity recognition using wearable sensors

In a worldwide scale, the population that suffers from muscular impairments, spinal cord injuries, or lower-limb amputations is estimated to be in millions, this population is in need for advanced rehabilitation devices to help them restore their lower-limb functionality. Accurate recognition of loco-motor activity is of great importance for improving rehabilitation devices that activate and take the correct action according to the current activity. However, this field of activity recognition lacks a comprehensive study on the suitability of some methods, algorithms or sensors over the others. In this paper, we carried out several experiments to reach for the best possible machine learning tools and sensor combinations or placements that can maximize off-line accuracy and minimize prediction time. Starting from a data set that combines sEMG and mechanical sensors placed bilaterally, after we specified 8 different loco-motor activities and extracted features that were used extensively in the literature, we found that EMG signals do not offer a significant improvement in accuracy unlike mechanical sensors which achieved alone an average of 97% accuracy in subject-dependent context. The optimal set of features for mechanical sensors and sEMG sensors were found to be time domain features and wavelet coefficients features respectively. The performance of KNN, LDA, QDA, LR, ANN, ETC and SVM models were compared; at which ETC was found to achieve the highest performance in terms of accuracy 92% and 97% in subject-dependent and subject-independent context respectively. LDA, LR and QDA outperformed the others in terms of time and model size measures.