Gait Cycle Detection by Markerless Tracking and Swing Phase Prediction of Paralyzed Leg for Neurostimulation Control*

Nerve stimulation is a technique used to restore walking ability in individuals who have suffered damage to their peripheral nerves. In order to reconstruct the walking motion, it is necessary to coordinate the stimulation with the gait cycle of the healthy parts of the patients. This study employed markerless tracking technology to demonstrate gait cycle identification and estimate the swing phase initiation in the paralyzed leg, utilizing gait periodicity as a basis. A Convolutional Neural Network (CNN) employing convolutional and pooling layers was applied to identify 22 anatomical landmarks on a rat’s body. The angular positions of the joints and the velocities of the toes were computed based on the body components identified by the CNN. The angles of the joints, velocities of the toes, and time series of both angles and velocities were employed as inputs to categorize the phase of walking. The initiation time of the swing phase in the paralyzed limb was subsequently determined based on the gait phase of the healthy legs. The estimated initiation of the swing phase was compared to the actual initiation of the swing phase in healthy rats. The proposed approach achieved an accuracy of $85 \%$ for detecting the swing phase initiation in the healthy foreleg and approximately $70 \%$ for the hindleg. Nevertheless, the accuracy of estimating the initiation of the swing phase in a paralyzed leg was measured at $60 \%$, indicating a need for improvement. This result came from the difficulty of precisely forecasting non-cyclical leg movements. Identifying non-cyclical moves will improve the estimation accuracy in the future.