This study proposes a method for the integration of a segmentation model with non-linear regression to construct weld-seam trajectories. First, the weld is precisely segmented using the YOLOv8 model. Then, its boundary pixels are extracted right after and termed SEG points. The 48, 24, and 16 SEG points are utilized as the input for regression models to estimate the weld seam center-line. Both linear and non-linear (quadratic) regression models are assessed using different types of weld images. The experimental results show that the average segmentation performance and training accuracy of the YOLO model are 94%. Additionally, both linear- and nonlinear regression models can estimate similar weld seam profiles. It is worth noting that the more the SEG points are utilized, the higher the accuracy is. However, if a number of SEG points increase, the processing time increases too. Therefore, 24 SEG points are enough for the precise estimation of the weld seam trajectory. It demonstrated that the proposed approach achieves a mean absolute trajectory error of 0.24 mm at 2.9 FPS for the end-to-end pipeline by using the Ampere GPU of NVIDIA Jetson Nano (input 416×416, FP16). Moreover, estimated weld seam trajectory is transformed into the absolute coordinate for...