Denoising of motion capture data with recurrent neural networks

Abstract

This thesis presents the problem of motion tracking and the concept of motion capture, and an overview of different methods for capturing a motion. This thesis provides a detailed description of marker-based optical motion capture and compares it with other methods for motion capture. In addition, introduction to neural networks and denoising of sequential data is provided. Practical section of the thesis entitles capturing approximately 30 minutes of motion data, training multiple recurrent neural networks to denoise it and comparing the results of multiple model types and a proprietary denoising software. Comparing is performed on three different metrics: mean square error between prediction and original sequence, mean difference in bone lengths between prediction and original sequence and mean difference of joint gradients between the prediction and the original sequence.