Marker-Less 3d Object Recognition and 6d Pose Estimation for Homogeneous Textureless Objects: An RGB-D Approach.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Object recognition and 6D pose estimation are vital for industrial automation, particularly for robotic manipulation.
• Consumer-grade RGB-D cameras offer a cost-effective solution for these tasks, even for small businesses.
• Existing methods struggle with textureless, homogeneous industrial parts in complex scenes.

Purpose of the Study:

• To develop a robust framework for simultaneous object recognition and 6D pose estimation from RGB-D data.
• To address the challenges posed by industrial parts, such as being textureless and homogeneous.
• To enable accurate robotic manipulation in industrial assembly lines.

Main Methods:

• A global approach is employed, first recognizing objects and Regions of Interest (ROIs) from RGB images.
• Object pose is subsequently estimated using depth information from RGB-D data.
• Classifiers based on Histogram of Oriented Gradient (HOG) features are used for recognition, followed by template matching on point clouds using surface normals and Fast Point Feature Histograms (FPFH) for pose estimation.

Main Results:

• The proposed framework demonstrates efficiency and accuracy in object recognition and 6D pose estimation.
• The system proves robust to variations in illumination and background.
• Effective performance was validated on challenging, textureless objects from the Tless dataset.

Conclusions:

• The developed framework offers a viable solution for object recognition and 6D pose estimation in industrial settings.
• The method's robustness makes it suitable for real-world applications with challenging object types.
• This work contributes to advancing robotic capabilities in industrial automation through improved perception.