Abstract: Light reflection, spatter, dust and other noises generated during automatic welding of thin plates blocked position information of weld, thus affecting identification and extraction of feature points. Therefore, a connected region algorithm was proposed to mark features of weld, and connected region algorithm was improved to extract feature points of weld and obtain its position information. Before image preprocessing, region of interest (ROI) method was used to segment image of laser stripe, which could filter out a large amount of noise such as arc and spatter. In the process of image preprocessing, binarization algorithms of median filter and maximum between-cluster variance were used to reduce interference noise near laser stripe, and laser stripe was separated from background to make characteristics of weld clearer and more obvious. After image preprocessing, laser stripe was marked by connected region algorithm, and position of connected region was determined by the improved algorithm, so as to identify weld feature points and obtain position information of weld feature points. The algorithm not only preserved edge information of laser stripe of weld, but also completed identification of weld features in a complex working environment. By comparing sheets’ actual gap width and gap width calculated by experiment, average error of the algorithm was within 0.067 mm, which met precision requirements in the industry, and it was suitable for weld tracking process of laser vision.