Research on Real-Time Welding Spot Extraction Algorithm for Welding Seam Tracking

The precision of weld point extraction is significantly influenced by disturbances such as arc light and splattering during welding processes. This paper proposes a real-time weld point extraction algorithm designed for weld seam tracking using binocular laser vision. Initially, to capture laser stripes during welding, we utilize the Yolov8 deep learning framework integrated with the lightweight HGNetv2 network backbone. The Neck section incorporates the FDPN (Focusing Diffusion Pyramid Network) to enhance feature detection accuracy. Introducing the InnerIoU bounding box loss in the loss function accelerates model convergence, resulting in a 1.9% improvement in stripe instance segmentation accuracy and achieving an inference speed of 172 FPS. After extracting laser stripes, image processing is applied to identify the coordinates of weld points from the left view. Spatial 3D coordinates of weld points are determined using left-right camera disparity, enabling real-time extraction of weld point 3D coordinates. Experimental validation confirms welding errors within 0.9 mm and a tracking speed of 33 FPS, meeting the real-time weld seam tracking requirements.