Laser stripe segmentation and centerline extraction based on 3D scanning imaging

Area of Science:

• Computer Vision
• Metrology
• Machine Learning

Background:

• High-precision 3D morphology measurement using structured light is hindered by ambient noise and illumination variations.
• Existing methods struggle to maintain accuracy under challenging environmental conditions.
• Accurate laser stripe center extraction is crucial for reliable 3D reconstruction.

Purpose of the Study:

• To develop a robust sub-pixel laser stripe center extraction method.
• To mitigate the impact of ambient noise and illumination inhomogeneity on structured light 3D measurements.
• To enhance the precision and repeatability of 3D morphology analysis.

Main Methods:

• A hybrid approach combining UNet deep learning for coarse segmentation and level set for fine segmentation of laser stripes.
• Utilizing UNet to obtain prior shape information for complex scenes.
• Improving the level set energy function with shape constraints and extracting stripe centers using a fused gray center of gravity method guided by the normal field.

Main Results:

• Experimental validation demonstrates a significant reduction in measurement errors.
• Average width error for point cloud data of workpieces with varying widths is less than 0.3 mm.
• Average repeatability extraction error is consistently below 0.2 mm.

Conclusions:

• The proposed sub-pixel extraction method effectively overcomes ambient noise and illumination inhomogeneity.
• The method achieves high precision and repeatability in structured light 3D morphology measurement.
• This technique offers a reliable solution for accurate 3D reconstruction in challenging environments.