國立高雄大學圖資館 |

Language: English

Back

Towards Human-Like Motion Prediction.

Carnegie Mellon University.

Towards Human-Like Motion Prediction.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Towards Human-Like Motion Prediction.
Author:	Gui, Liangyan.
Published:	Ann Arbor : ProQuest Dissertations & Theses, 2019
Description:	142 p.
Notes:	Source: Dissertations Abstracts International, Volume: 80-09, Section: B.
Notes:	Publisher info.: Dissertation/Thesis.
Notes:	Advisor: Moura, Jose M. F.
Contained By:	Dissertations Abstracts International80-09B.
Subject:	Robotics.
Online resource:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=13807187
ISBN:	9780438974500

Towards Human-Like Motion Prediction.
Gui, Liangyan.

Towards Human-Like Motion Prediction. - Ann Arbor : ProQuest Dissertations & Theses, 2019 - 142 p.

Source: Dissertations Abstracts International, Volume: 80-09, Section: B.

Thesis (Ph.D.)--Carnegie Mellon University, 2019.

This item is not available from ProQuest Dissertations & Theses.

In this dissertation, we address predictive learning in the context of 3D human motion prediction — forecasting human motion from hundreds of milliseconds to a few seconds given a historical skeleton sequence. This ability is missing in modern artificial intelligent systems. State-of-the-art deep learning based approaches typically formulate the task as a sequence-to-sequence problem and solve it by using recurrent encoder-decoder neural networks. Despite notable successes, these existing approaches suffer from prediction discontinuity, rely on extensive annotated motion capture data, are brittle to novel actions, and do not perform well in longer time horizons due to error accumulation and uncertainty. We focus on human-like motion prediction so that the predicted sequences are more plausible, realistic, and temporally coherent with past sequences in both short-term and long-term situations for a variety of actions. Our key insight is to exploit the rich yet implicit structural dependencies and regularities within motion sequences, including geometric, temporal, model parameter, and contextual structures without any additional supervision. In this spirit, we tackle key technical challenges and explore complementary perspectives. We integrate these perspectives into a deep learning based prediction framework, and leverage group theory, adversarial learning, meta-learning, and attention mechanism to acquire the desired structural information. We start by addressing the fidelity and continuity in deterministic prediction. We incorporate local geometric structure constraints through a frame-wise geodesic loss on a Lie group. Further, we simultaneously validate the sequence-level plausibility of the prediction and its coherence with the input sequence by introducing two global recurrent discriminators together with adversarial learning. Next, we consider a crucial yet under-explored issue, namely, the small sample size problem. We deal with, which we believe for the first time, the few-shot prediction and propose a general proactive and adaptive meta-learning framework that enables rapid generation of a task-specific prediction model for a novel action from few annotated motion sequences. A third problem we tackle is the uncertainty and stochasticity inherent in long-term prediction. We predict multiple plausible future sequences through multiple choice learning over an ensemble of attention-based predictors. Each of the predictors produces a context-dependent prediction by paying attention to different spatial and temporal evolutions of the past motion. We benchmark our results with the largest-scale, widely-used Human 3.6M dataset. We show that our approaches significantly outperform current state-of-the-art results under various criteria. Finally, we deploy our prediction models into practical systems, such as (1) teaching a humanoid robot "Pepper" to interact with a human by predicting and mimicking how the human moves or acts, and (2) synthesizing and animating human motion with a virtual human body "Adam" on Unity, a graphics rendering platform.

ISBN: 9780438974500Subjects--Topical Terms:

181952
Robotics.

Towards Human-Like Motion Prediction.
LDR:04246nmm a2200349 4500 001 570711
005 20200514111943.5
008 200901s2019 ||||||||||||||||| ||eng d
020 $a 9780438974500
035 $a (MiAaPQ)AAI13807187
035 $a (MiAaPQ)cmu:10362
035 $a AAI13807187
040 $a MiAaPQ $c MiAaPQ
100 1 $a Gui, Liangyan. $3 857360
245 1 0 $a Towards Human-Like Motion Prediction.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2019
300 $a 142 p.
500 $a Source: Dissertations Abstracts International, Volume: 80-09, Section: B.
500 $a Publisher info.: Dissertation/Thesis.
500 $a Advisor: Moura, Jose M. F.
502 $a Thesis (Ph.D.)--Carnegie Mellon University, 2019.
506 $a This item is not available from ProQuest Dissertations & Theses.
506 $a This item must not be sold to any third party vendors.
520 $a In this dissertation, we address predictive learning in the context of 3D human motion prediction — forecasting human motion from hundreds of milliseconds to a few seconds given a historical skeleton sequence. This ability is missing in modern artificial intelligent systems. State-of-the-art deep learning based approaches typically formulate the task as a sequence-to-sequence problem and solve it by using recurrent encoder-decoder neural networks. Despite notable successes, these existing approaches suffer from prediction discontinuity, rely on extensive annotated motion capture data, are brittle to novel actions, and do not perform well in longer time horizons due to error accumulation and uncertainty. We focus on human-like motion prediction so that the predicted sequences are more plausible, realistic, and temporally coherent with past sequences in both short-term and long-term situations for a variety of actions. Our key insight is to exploit the rich yet implicit structural dependencies and regularities within motion sequences, including geometric, temporal, model parameter, and contextual structures without any additional supervision. In this spirit, we tackle key technical challenges and explore complementary perspectives. We integrate these perspectives into a deep learning based prediction framework, and leverage group theory, adversarial learning, meta-learning, and attention mechanism to acquire the desired structural information. We start by addressing the fidelity and continuity in deterministic prediction. We incorporate local geometric structure constraints through a frame-wise geodesic loss on a Lie group. Further, we simultaneously validate the sequence-level plausibility of the prediction and its coherence with the input sequence by introducing two global recurrent discriminators together with adversarial learning. Next, we consider a crucial yet under-explored issue, namely, the small sample size problem. We deal with, which we believe for the first time, the few-shot prediction and propose a general proactive and adaptive meta-learning framework that enables rapid generation of a task-specific prediction model for a novel action from few annotated motion sequences. A third problem we tackle is the uncertainty and stochasticity inherent in long-term prediction. We predict multiple plausible future sequences through multiple choice learning over an ensemble of attention-based predictors. Each of the predictors produces a context-dependent prediction by paying attention to different spatial and temporal evolutions of the past motion. We benchmark our results with the largest-scale, widely-used Human 3.6M dataset. We show that our approaches significantly outperform current state-of-the-art results under various criteria. Finally, we deploy our prediction models into practical systems, such as (1) teaching a humanoid robot "Pepper" to interact with a human by predicting and mimicking how the human moves or acts, and (2) synthesizing and animating human motion with a virtual human body "Adam" on Unity, a graphics rendering platform.
590 $a School code: 0041.
650 4 $a Robotics. $3 181952
650 4 $a Artificial intelligence. $3 194058
650 4 $a Computer science. $3 199325
690 $a 0771
690 $a 0800
690 $a 0984
710 2 $a Carnegie Mellon University. $b Electrical and Computer Engineering. $3 660346
773 0 $t Dissertations Abstracts International $g 80-09B.
790 $a 0041
791 $a Ph.D.
792 $a 2019
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=13807187