Abstract: In view of small working space between cabins and difficulty of large-scale automation equipment to complete automatic welding in the cabin, a strategy based on binocular vision to obtain three-dimensional information of weld path was proposed. After obtaining basic parameters of binocular system based on Zhang Zhengyou’s calibration principle, C++ and OpenCV were used to write programs such as binarization of adaptive threshold, improved Sobel contour extraction operator and non-continuous pixel screening, and extract a clear, low-noise image of center contour of the right angle weld. Based on BM feature point matching algorithm and pixel scanning method, a three-dimensional information data set of continuous feature points on weld outline was obtained, and a three-dimensional right-angle weld path was generated after fitting with Origin drawing software. In order to test accuracy of binocular system ranging, a universal joint combination module of sliding rails with adjustable angles was designed to collect weld images from different shooting angles and heights, and to identify distance of equidistant feature points set on weld. The experimental results showed that when shooting deflection angle was within 30° or shooting height was within 150～190 mm, ranging deviation could be controlled within 2 mm, which basically satisfied the precision and stability requirements of welding, and provided a data basis for automatic tracking process of welding.Highlights: (1) Binocular visual recognition ranging technology was applied to identification of center path of rectangular fillet welds in the small cabin lattice, which provided tracking data basis for the automatic welding process.(2) Slide rail universal junction module was designed, which confirmed good robustness and recognition accuracy of binocular visual recognition system at different angles and heights.