Human motion enhancement through deep learning

Human motion analysis is a rapidly developing field of computer vision with a wide variety of applications including biomedical study, athletic observation, computer interfaces, entertainment media development, and video classification. The goal of human motion analysis is to model the human body in a concise form, generally as a skeletal model. I propose a new deep learning method for enhancement of human motion data through the use of non-linear Kalman filtering inside of the latent space of a convolutional autoencoder trained on high quality optical Mocap human motion data. This method takes advantage of a learned human motion manifold as well as preserving the original motion kinematics through the use of Kalman Filtered data. In addition, I propose a new simulated environment designed to generate human depth data in a multi-view environment. This system presents the advantage of a $360^{\circ}$ pose capture space and is used to perform transfer learning on a voxel-based deep neural network. This work exemplifies the usefulness of simulated data when real-world data is scarce. In completing this work I have generated two new datasets, the extended MHAD dataset, based on real-world depth data, and the Human Motion Depth Dataset (HuMoD), both of which will be made available to the public.