| Resumo: | Real-time locomotion mode recognition can potentially be applied in the gait analysis as a diagnostic tool or a strategy to control the robotic motion. This research aimed the development of an automatic, accurate and time-effective tool to recognize, in real-time, the locomotion mode that is being performed by a humanoid robot. The proposed strategy should also be general to different walkers and walking conditions. For these purposes, we designed a strategy to identify, in an offline phase, the suitable features and classification models for the real-time recognition. We explored several classification models based on two machine learning approaches using the features previously selected by principal component analysis and genetic algorithm (GA). The validation was carried out for distinct walking directions and speeds of DARwIn-OP. The offline analysis suggests that the most skilled models are the ones created by weighted k-nearest neighbors (KNN), fine KNN, and cubic support vector machine using 2 features selected by GA. Results from the real-time implementation highlight that weighted KNN exhibits a higher recognition performance (accuracy > 99.15%) and a lower elapsed time in the recognition process (89 ms) comparatively to the state-of-the-art. The proposed recognition tool showed to be cost-effective, and highly accurate for the real-time gait analysis at different walking conditions.
|
|---|