Motion Capture in Mixed-Reality Applications: A Deep Denoising Approach

Motion capture is a fundamental technique in the development of video games and in film production to animate a virtual character based on the movements of an actor, creating more realistic animations in a short amount of time. One of the ways to obtain this movement from an actor is to capture the motion of the player through an optical sensor to interact with the virtual world. However, during movement some parts of the human body can be occluded by others and there can be noise caused by difficulties in sensor capture, reducing the user experience. This work presents a solution to correct the motion capture errors from the Microsoft Kinect sensor or similar through a deep neural network (DNN) trained with a pre-processed dataset of poses offered by Carnegie Mellon University (CMU) Graphics Lab. A temporal filter is implemented to smooth the movement, given by a set of poses returned by the deep neural network. This system is implemented in Python with the TensorFlow application programming interface (API), which supports the machine learning techniques and the Unity game engine to visualize and interact with the obtained skeletons. The results are evaluated using the mean absolute error (MAE) metric where ground truth is available and with the feedback of 12 participants through a questionnaire for the Kinect data.