Lower limb gait activity recognition using Inertial Measurement Units for rehabilitation robotics

In this paper, The authors considered a human lower limb gait activity recognition algorithm, using an IMU sensory network consisting of 4 IMUs distributed to the lower limb. The proposed algorithm depends on Random Forest for classification and a Hybrid Mutual Information and Genetic Algorithm (HMIGA) as a features selection technique. HMIGA selects the most distinguishing features from Discrete Wavelet Coefficient (DWT) features and other statistical and physical (self designed) features. The proposed algorithm is compared with Support Vector Machine (SVM) to classify 5 activities and the results are presented on 6 subjects with 2% average error rate with 1.9% superiority on SVM. Moreover, HMIGA as a feature selector is compared to the traditional feature selectors and DWT as a feature also compared to statistical and physical features, showing their influence on the activity recognition process. Finally, the most important features selected by HMIGA are presented, proving the important role of the shank´s sensor on the recognition process, where almost 50% of the selected features are from the shank sensor.