Prediction and Estimation of Human Motion using Generative-Adversarial Network
Open Access
Author:
Subbakrishna Adishesha, Amogh
Graduate Program:
Electrical Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
July 10, 2018
Committee Members:
Robert Collins, Thesis Advisor/Co-Advisor David Jonathan Miller, Committee Member Mehdi Kiani, Committee Member
Keywords:
GANs LSTM Motion Prediction Deep Learning Neural Networks AI Prediction Error
Abstract:
Prediction of the human motion model has been an intrinsic part of several applications
over diverse fields like gaming, augmented reality and cinematic graphics. The ability to estimate
motion, ahead of time, helps robots predict human action and thus reduce time to react
effectively. In real time applications such as pedestrian motion prediction, the availability of long
motion sequences at test time is rare. In this work, we propose a new architecture to predictively
model human motion partially from noise. We utilize the data synthesizing ability of Generative
Adversarial Networks(GANs) to provide artificial motion frames that help in prediction of the
motion sequence in an LSTM-RNN framework. The well proven Recurrent Neural Network is
used as a discriminator in training a weaker LSTM generator that we later exploit in creating
ground truth like data from randomly sampled frames with mean pose and added noise. Pivoting
on the evaluation metrics used in latest works, we discuss the recent motion prediction techniques
and compare the results. We also evaluate the training procedures, input requirements and
complexity of the structures, thus illustrating the simplicity and accuracy of a GAN based input
reduction model.
