Abstract: Automatic extraction of weld feature points is the key to autonomous weld paths planning for robot. Conventional welding usually requires teaching welding intermediate points, which usually needs to be re-teached when the workpiece is changed or there is an error in the assembly, resulting in low welding efficiency. In this paper, we propose a method for automatic extraction of welding feature points based on a monocular camera, bilateral filtering is used to smooth the image; then Sobel operator is applied to calculate the gradient of the pixels and the average gray value of the image is used as the threshold to binary pixels; then a closed operation is used to fill the fine holes and connects the intermittent fine contour lines;then an improved contour tracking operator is used to eliminate residual noise on the image;then the feature points are taken from image and converted from the pixel coordinate system to the user coordinate system by camera calibration.Different shapes of flat sheet butt weld workpieces are chosen for feature point extraction experiments, by comparing the difference between the feature points and the weld bevel center points obtained by laser sensor in the y-axis direction of the user coordinate system,the average errors are 0.45 mm, 0.23 mm, 0.42 mm for hybrid, linear and curve welds respectively,which indicating that the method can automatically identify weld from the image and extract the weld feature points, furthermore, eliminate teaching of the welding system.