Aiming at the problem of multiple extraction or incomplete extraction of the center position of the light strip caused by overexposure and underexposure of the fringe image in the three-dimensional measurement of complex objects, a light strip center line extraction method based on deep learning semantic segmentation technology is proposed. The method uses the improved UNet++ network for semantic segmentation, roughly extracts the center area of the light strip, and obtains the light strip center line with one to two pixel width, Then the sub-pixel center is accurately extracted by the gray center of gravity method. Experiments show that this method can effectively overcome the adverse effects caused by uneven exposure of light bar images, and accurately extract the complete and smooth subpixel light bar centerline of complex surface objects.