Long short-term memory assisted FIR filter for tracking human upper limb motion

To better correct the spatial trajectory of human motion, this paper proposes a Long Short-Term Memory (LSTM)-assisted correction algorithm for the Finite Impulse Response (FIR) filter, aimed at improving the accuracy of human joint motion capture based on inertial sensors. In the proposed algorithm, the forward data is obtained from the backward-updated data smoothed by the Rauch-Tung-Striebel (RTS) smoother and compared with the forward-predicted data from the FIR. The resulting error is input into the pre-trained LSTM model to predict the error data, and the backward data at that moment is updated, making the positional information more accurate. This calculation shows significant effectiveness when estimating the elbow data. In the experiment, the data obtained using the LSTM-assisted FIR filter showed a 35.83% improvement compared to the data obtained with the LSTM-assisted Kalman Filter.