Aiming to address the issues of poor stability and low accuracy of laser stripe centerline extraction algorithm in complex environments, a novel centerline extraction method based on improved U²-Net is proposed. Firstly, TSA (transformer-self-attention) and TCA (transformer-cross-attention) modules are added to the U²-Net network to improve the feature extraction ability of the model, achieve accurate pixel-level segmentation, effectively remove noise and glitches in the image, and provide high-quality image sources for subsequent centerline extraction. Secondly, according to the characteristics of the application scenario, the traditional Steger method is improved to complete the high-precision extraction of the centerline of the laser stripe. Finally, the reliability value evaluation mechanism is used to analyze the accuracy of the center point of the light stripe. Experimental results show that compared with other mainstream semantic segmentation networks, the improved U²-Net proposed in this paper has higher extraction accuracy and better anti-noise performance, and the reliability value of the extracted pixel center point on this basis is higher, reaching 1.9 times that of the traditional Steger algorithm, which meets the needs of high-precision industrial measurement.