Real-time Walking Speed Estimation by Involving Phase Variables based on Deep Learning

Lower limb prothesis is neccessary in daily avtivities for unilateral transfemoral amputees and gait analysis plays an important role in user-friendly knee-ankle prothesis research. Walking speed is recognized as an important feature that influences joint trajectories, the estimation of which is vitual to guarantee walking fluency and symmetry. In this paper, walking speed is estimated in real-time by involving phase variable and radius as the inputs of deep learning network. Virtual Constraint control algorithm is used to generate prosthetic joints trajectories as functions of phase variable through the entire gait by collecting thigh angle from an Inerial Measurement Unit (IMU) sensor mounted on the thigh. Besides, Long Short-Term Memory (LSTM) network is applied to train the estimation model. In consideration of the tight relationship between walking speed and phase plane curves, this research utilizes horizontal and vertical coordinates, variable and phase radius as the inputs of deep learning network instead of directly using raw sensors data. The simulation experiments proved that this algorithm could estimate the speed effectively, and be more accuate than using raw sensors data.