Related Experiment Video
Updated: May 13, 2026
09:41
Estimation of Contact Regions Between Hands and Objects During Human Multi-Digit Grasping
Published on: April 21, 2023
Tracking people's hands and feet using mixed network AND/OR search
Vlad I Morariu1, David Harwood, Larry S Davis
1Department of Computer Science, Institute for Advanced Computer Studies, University of Maryland, College Park, MD 20742, USA. morariu@cs.umd.edu
Summary
This study introduces a novel framework for tracking multiple people's hands and feet in 2D videos, even with occlusions. It uses an efficient optimization method to achieve accurate, real-time human pose estimation from a single camera.
Area of Science:
- Computer Vision
- Machine Learning
- Robotics
Background:
- Accurate multi-person pose estimation from single camera views is challenging due to occlusions and complex interactions.
- Existing methods often require multiple views, high-resolution images, or extensive training data.
Purpose of the Study:
- To develop an efficient framework for detecting and tracking multiple interacting humans' hands and feet in 2D from a single camera.
- To overcome limitations of existing methods by requiring fewer constraints and handling occlusions effectively.
Main Methods:
- A part-based formulation emphasizing extremities for globally optimal solutions in each frame.
- An efficient optimization scheme utilizing AND/OR Branch-and-Bound with lazy evaluation and cost-sensitive bounds.
- Leveraging mixed probabilistic and deterministic networks and their AND/OR search space.
Main Results:
- State-of-the-art performance in single-person tracking on the HumanEva dataset.
- Demonstrated robustness in multi-person scenarios with partial and full occlusions and fast motion in real-world datasets.
- Successful detection and tracking of heads, hands, and feet through occlusions.
Conclusions:
- The proposed framework offers an efficient and accurate solution for multi-person pose estimation from single-view 2D data.
- The method is robust to occlusions and complex interactions, requiring minimal constraints.
- This approach advances human motion analysis in unconstrained environments.
