Motion Capture Data Denoising Based on LSTNet Autoencoder

Abstract

This paper proposes a novel deep learning-based optical motion capture denoising model encoder-LSTNet- decoder (ELD). ELD uses an autoencoder for manifold learning and decoder to remove jitter noise and missing noise effectively. It uses recurrent units in LSTNet to effectively obtain the spatial-temporal information of motion sequences, especially the periodic long-term and short-term features. In the denoising procedure, the kinetical characteristics of the motion are also considered so that the reconstructed deviation is smaller and can more accurately reflect the real action. We simulated ELD with the CMU database and compared it with the art-of-state methods. The experiment shows that ELD is a very effective denoising technique with lower reconstruction error, stronger robustness, and shorter running time.