Related Experiment Video
Updated: Dec 1, 2025
09:41
Estimation of Contact Regions Between Hands and Objects During Human Multi-Digit Grasping
Published on: April 21, 2023
2.0K
Robust 3D Hand Detection from a Single RGB-D Image in Unconstrained Environments
Chi Xu1,2,3, Jun Zhou1,2, Wendi Cai1,2
1School of Automation, China University of Geosciences, Wuhan 430074, China.
Sensors (Basel, Switzerland)
|November 11, 2020
Summary
This study introduces a robust 3D hand detection method using adaptive RGB-D channel fusion. The approach significantly enhances accuracy and stability in unconstrained environments, outperforming existing methods in challenging lighting conditions.
Area of Science:
- Computer Vision
- Robotics
- Human-Computer Interaction
Background:
- 3D hand detection from RGB-D images is vital for numerous applications.
- Unconstrained environments pose challenges due to factors like lighting variations affecting RGB and depth (D) channels.
Purpose of the Study:
- To develop a robust and accurate 3D hand detection approach by adaptively fusing RGB and D features.
- To improve performance in unconstrained environments and under varying lighting conditions.
Main Methods:
- A novel approach adaptively fuses complementary features from RGB and D channels.
- A cascaded network first detects 2D hand bounding boxes, then estimates 3D locations.
- A new RGB-D hand detection dataset for unconstrained environments was created.
Main Results:
- The proposed RGB-D fusion method achieved 74.1% accuracy, a significant improvement over RGB-based detectors (69.1%).
- The D-channel proved crucial for detection in unconstrained settings.
- The fusion approach demonstrated superior robustness in extreme lighting (dark: 62.5%, back-light: 65.9%) compared to single-channel methods.
Conclusions:
- Adaptive RGB-D feature fusion enhances 3D hand detection robustness and accuracy.
- The method significantly outperforms existing single-channel approaches, especially in challenging and variable conditions.
