IMU-based continuous prediction of human lower limb joint angles

Predicting human motion intent is crucial for exoskeleton robots to provide effective assistance. In recent years, many deep learning-based predicting methods have been shown to provide effective prediction of human movement intentions. On the other hand, current researchers often ignore the cyclic and rhythmic nature of human walking. We propose an Adaptive Multi-Oscillator Kalman Filter Prediction (AMOKFP) model that uses kinematic data to predict human movement states over a short period of time in the future. Our approach consists of three stages: gait feature learning, parameter prediction and state prediction optimization. In the first stage, we decompose and learn human motion features. In the second stage, we perform numerical prediction on the learning results from the first phase. In the third stage, we perform optimal estimation on the prediction results from the second phase. The simulation results show that the model has a good estimation performance and the prediction error is stable at 0.13°. The AMOKFP model proposed in this paper can be applied to the control system of lower limb-assisted exoskeleton robots as a part of state perception and involved in the planning of exoskeleton assisting moments.