Estimation of Foot Ground Contact from Monocular Video for Physically Plausible Gait Animation
- 주제(키워드) foot-ground contact detection , video processing , motion reconstruction
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 유리
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035676
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Foot-ground contact information plays a crucial role in character animation and gait analysis, as it helps in accurately simulating realistic movement patterns and understanding the biomechanics of walking. Existing motion datasets do not explicitly include foot-ground contact information, requiring separate computation or manual annotation. Obtaining accurate foot-ground contact information typically requires additional sensors such as pressure mats or force plates. Without such devices, estimating contact becomes a highly challenging task. We propose ContactVision, a deep learning framework that detects heel and toe contact states directly from video. Our network is trained in a supervised manner using contact labels derived from motion capture data via ground reaction force estimation. This enables training on existing datasets without the need for additional hardware. We demonstrate the utility of our contact detection network in two downstream tasks: gait motion reconstruction and gait analysis. For animation, we incorporate predicted contact labels into a reinforcement learning framework with a two-segment foot model, enabling realistic foot articulation behavior. For analysis, we estimate clinically relevant gait parameters such as double and single support times, and validate the accuracy against pressure sensor mat data and prior video-based methods. Our results show competitive performance in both animation and analysis settings.
more목차
1. Introduction 1
2. Related Work 5
2.1 Foot Contact Acquisition 5
2.2 Applications of Foot Contact Information 7
3. Overview 9
4. Training Data Preparation 11
4.1 Data Specification 11
4.2 Keypoint Extraction 12
4.3 Contact Labeling 13
5. Contact Detection Network 16
5.1 Network Architecture 16
5.2 Network Training 17
6. Results 19
6.1 Implementation Details 19
6.2 Performance Evaluation 20
6.3 Application 1: Motion Reconstruction 22
6.4 Application 2: Gait Analysis 33
6.4.1 Gait Phase Classification 33
6.4.2 Gait Parameters Extraction 35
6.4.3 Evaluation 36
7. Discussion 38
Bibliography 40
Appendix 47
A Marker Mapping 47
B Detailed Gait Parameter Measurements 50
C Details of Reward Terms Used in Motion Reconstruction 54
