Simultaneous Recognition and Relative Pose Estimation of 3D Objects Using 4D Orthonormal Moments

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Three-dimensional (3D) object recognition and pose estimation are critical research areas with diverse applications.
• Existing algorithms often address these tasks separately or jointly with varying complexities.
• A need exists for efficient and accurate methods capable of concurrent 3D object recognition and pose estimation.

Purpose of the Study:

• To introduce a novel algorithm for simultaneous 3D object recognition and pose estimation.
• To demonstrate the analytical capabilities of the proposed method without reliance on experimental work.
• To validate the algorithm's performance on textureless, low-resolution projections under various rotations and real-world conditions.

Main Methods:

• A four-dimensional (4D) tensor is defined to organize 3D object projections from multiple viewpoints.
• The 4D tensor is represented by 4D orthonormal moments, pre-computed into a matrix.
• Recognition and pose estimation are achieved by solving a linear least squares problem using the matrix and 2D moments of observed projections.

Main Results:

• The method analytically proves its efficacy for both 3D object recognition and pose estimation.
• The algorithm demonstrates computational simplicity and high performance, suitable for real-time applications.
• Experiments involving yaw and pitch rotations, as well as real-world internet images, show encouraging results comparable to state-of-the-art methods.

Conclusions:

• The proposed algorithm offers a computationally efficient and effective solution for simultaneous 3D object recognition and pose estimation.
• Its ability to handle textureless, low-resolution projections and its strong performance validate its applicability in demanding scenarios.
• The method represents a significant advancement in the field, providing a robust approach for complex 3D vision tasks.