Proceedings Article | 5 December 2024
KEYWORDS: Point clouds, Calibration, Cameras, Education and training, Image registration, Detection and tracking algorithms, Eigenvectors, Data processing, Attenuation, 3D image processing
Manual wheel size inspection faces problems such as limited accuracy, low efficiency, and poor consistency. Machine vision technology can improve measurement accuracy, efficiency and consistency through high-resolution cameras and advanced image processing algorithms. However, due to factors such as difficult railside deployment, limited camera field of view and occlusion, it becomes difficult to obtain whole wheel data, thus affecting wheel size measurements. To address this problem, this paper proposes an Iterative Closest Point (ICP) algorithm improved by sampling consistency and distance threshold decay based on the Random Sample Consensus (RANSAC) algorithm. Then, by adding calibration blocks and combining this algorithm to complete the point cloud registration of the whole wheel. Finally, based on the whole wheel data, the point set of the base point rounding and the wheel profile is extracted to complete the calculation of parameters such as wheel diameter, rim height, and rim thickness. The experimental results show that the registration algorithm in this paper improves the quality of registration by 24.311% on average compared with the original one. The measurement errors of wheel diameter, rim height, and rim thickness are only 0.5844mm, 0.2437mm, and 0.5484mm, which account for 0.0640%, 0.8660%, and 1.8124% of the dimensions, respectively, and satisfy the requirements of wheel detection accuracy for trains.
