A fast monocular 6D pose estimation method for textureless objects based on perceptual hashing and template matching.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Object pose estimation is crucial for industrial applications but often requires expensive hardware and computationally intensive deep learning models.
• Current methods demand significant resources (3D cameras, GPUs) and large labeled datasets, increasing costs and development time.

Purpose of the Study:

• To develop a fast, robust, and cost-effective object pose estimation algorithm.
• To reduce computational demands and hardware requirements for 6D pose estimation.
• To enable efficient pose estimation on resource-constrained devices.

Main Methods:

• A template-based matching algorithm utilizing a novel perceptual hashing method for binary images.
• Automatic preselection of relevant templates to reduce inference time.
• Benchmarking on automotive parts and public datasets, including experiments on synthetic data and evaluations under occlusion and noise.

Main Results:

• The proposed method achieves high accuracy and robustness on cost-effective hardware without GPU support.
• Demonstrates a superior trade-off between accuracy and computation time compared to previous methods.
• Achieves specific performance metrics (e.g., $$ rotation error, 14% translation error, $$ processing time) on an NVIDIA AGX Orin device.

Conclusions:

• The algorithm offers an efficient solution for object pose estimation, particularly for applications prioritizing hardware cost and power efficiency.
• It provides a favorable accuracy-processing time balance, making it suitable for resource-constrained environments.
• The method shows robustness to partial occlusions and noisy inputs, with potential for broad industrial adoption.