Abstract: Against the backdrop of global aging,knee exoskeletons are widely used in the maintenance of knee joint health and rehabilitation training for the elderly.Knee exoskeletons often employ embedded devices for the recognition of lower limb motion states,which requires finding a balance between the selection and layout of sensors,and the accuracy and computational complexity of algorithms.Therefore,this paper studies a human motion recognition algorithm suitable for knee exoskeletons,which uses two inertial measurement units(IMUs) on the thigh and lower leg to collect lower limb motion data.The recognition method includes three steps:feature combination,feature selection,and motion state recognition.It optimizes feature representation through cross-method feature combination.The improved One-vs.-Rest(OvR) strategy is applied to address motion recognition issues,within which a fusion algorithm combining Relief with Pearson correlation coefficients and a machine learning backward selection method is used for feature selection to reduce computational complexity,and historical information along with other state data are integrated into the model training to further enhance accuracy.The model classifies six types of daily motion states with an accuracy rate of up to 97.76%.Experimental results verify that the proposed algorithm can accurately and quickly recognize lower limb motion states under the limitation of a limited number of sensors,providing an accurate and low computational requirement solution for precise detection and real-time control of knee exoskeletons.

Key words: Motion recognition, IMU sensors, Feature combination, Machine learning, OvR cascade framework