Main article

Yingzi Wei
School of Mechanical Engineering, Shenyang Ligong University, Shenyang 110159, Liaoning, China
Chao Hu*
School of Mechanical Engineering, Shenyang Ligong University, Shenyang 110159, Liaoning, China
huchao@sylu.edu.cn
Xin Tao
School of Mechanical Engineering, Shenyang Ligong University, Shenyang 110159, Liaoning, China
Hang Zhang
School of Mechanical Engineering, Shenyang Ligong University, Shenyang 110159, Liaoning, China

Abstract

Flexible manufacturing loading systems for irregular workpieces require robust and computationally efficient pose estimation to achieve reliable robotic grasping without the need for costly pre-oriented feeding fixtures. Deep learning-based pose detection methods, while achieving high accuracy on benchmark datasets, impose prohibitive data collection burdens and exhibit limited robustness in the cluttered, variable-illumination environments of industrial material feeding systems. This paper proposes a monocular vision pose estimation framework that constructs deterministic line feature invariants (LFI) from single workpiece images to achieve robust 6-DOF pose estimation without training data. The method extracts line segment features from workpiece contours, constructs invariant descriptors relative to the workpiece orientation axis, and establishes a local logarithmic polar coordinate system for rotation-invariant representation. A rule-based knowledge system classifies workpiece types and estimates face-upward/face-downward orientation using normalized feature evidence, with an adaptive confidence tolerance function determining the final pose estimate to maximize pass-through volume. Experiments on three types of irregular castings and forgings demonstrate face-upward pose estimation accuracies of 98.8%, 96.3%, and 97.8% with mean angle errors of 0.022%, 0.025%, and 0.026% respectively. Processing time of 12.3 ms per image substantially outperforms CNN-based alternatives (487--612 ms), confirming suitability for real-time industrial deployment. Robustness evaluation under Gaussian image noise demonstrates that LFI maintains above 91% accuracy at noise level sigma = 30 pixels, compared to 65% for CNN and 48% for Hough transform baselines, establishing the framework as a practical solution for flexible loading systems in industrial information integration contexts.

Article details

How to Cite

Wei, Y., Hu, C., Tao, X., & Zhang, H. (2023). Pose Estimation of Dispersed Irregular Workpieces Using Monocular Vision with Line Feature Invariants and Logarithmic Polar Coordinate System. Journal of Intelligent Industrial Convergence, 3(3), 1-10. https://doi.org/10.63646/jiic.2023.030301