A Real-Time Grasp Detection Network Based on Multi-scale RGB-D Fusion

Object grasping is a fundamental task for robots, and grasp detection plays a crucial role in enabling robots to recognize and grasp objects. In grasp detection, RGB and depth (RGB-D) data are widely used and employing the fusion of RGB and depth data often performs better than only using single-mod...

Full description

Saved in:
Bibliographic Details
Published in2024 International Joint Conference on Neural Networks (IJCNN) pp. 1 - 8
Main Authors Liu, Dingxi, Feng, Xiaodong
Format Conference Proceeding
LanguageEnglish
Published IEEE 30.06.2024
Subjects
Accuracy
deep learning
Fuses
grasp detection
Grasping
multi-scale RGB-D fusion
Neural networks
Real-time systems
robot
Robots
Online AccessGet full text

Cover

More Information
Summary:Object grasping is a fundamental task for robots, and grasp detection plays a crucial role in enabling robots to recognize and grasp objects. In grasp detection, RGB and depth (RGB-D) data are widely used and employing the fusion of RGB and depth data often performs better than only using single-modality data alone. Currently the mainstream methods use early fusion strategy to fuse RGB and depth data. These methods struggle to fully exploit the information from both modalities, which degrades the performance of grasp detection. In this paper, a multi-scale RGB-D fusion grasp detection network is proposed to address this problem. Experiments on the public Jacquard dataset indicate that our method excels in detection accuracy and generalization without sacrificing efficiency, our method achieves 95.2% accuracy in IW split and 94.4% accuracy in OW split with an efficient average inference speed of 6 ms per image.
ISSN:2161-4407
DOI:10.1109/IJCNN60899.2024.10651134