A Neural-Dynamic Architecture for Concurrent Estimation of Object Pose and Identity

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Object recognition and pose estimation are crucial for human-robot interaction and manipulation.
• Existing methods often require extensive training data or struggle with dynamic environments.

Purpose of the Study:

• To develop a neurally inspired architecture for efficient object instance learning and pose estimation.
• To enable real-time object handling and interaction in dynamic tabletop scenarios.

Main Methods:

• A novel architecture stores features (color and edge histograms) from single object views.
• Real-time processing utilizes neural dynamics for feature extraction and matching.
• Pose estimation is performed concurrently with object recognition by aligning learned views.

Main Results:

• Achieved 87.2% recognition rate with single training views for 30 objects, alongside precise pose estimation.
• Demonstrated successful tracking of moving objects and segmentation of visual scenes.
• Attained 91.1% recognition on the COIL-100 dataset using four training views per object.

Conclusions:

• The proposed system offers robust and efficient object recognition and pose estimation capabilities.
• It effectively handles dynamic scenes and limited training data, suitable for real-world applications.
• The neurally inspired approach shows promise for advanced robotic perception and interaction.