Predicting missing markers in mocap data using LSTNet

Abstract

Aiming at the noise caused by missing marker data in optical human motion capture, an improved LSTNet neural network model was proposed in this paper, which decomposed the noise prediction into linear part and nonlinear part. In the nonlinear part, convolutional neural network and recurrent neural network are used to deal with periodic prediction, and LSTM is used to replace the gated recurrent unit GRU to enhance memory function. The linear part uses autoregressive models to deal with aperiodic predictions. Finally, the loss function based on the position of markers is constructed to improve the prediction accuracy. The simulation results show that the proposed denoising technique can obtain lower reconstruction error and strong robustness, and the reconstructed motion sequence is very close to the real motion sequence.