Robust joint recognition with structured-light vision sensing

A model-based method is proposed for recognition of weld joints composed of line segment features.Shape features of the structured-light stripe centerline are extracted sequentially by calculating the peak mass position of gray distribution for the image columns and fitting the piecewise line segments using the least squares method, then stripe profile is described qualitatively by predefined primitive labels, and lastly a hypothesize-verification based model matching algorithm is implemented to recognize the joint.During the matching, undetected portion of joint profile, due to optical noises such as welding spatters, secondary reflection, tack weld, and other unexpected imaged sources, can be recovered to ensure the correct recognition results by merging profile's primitives and reconstructing the relations with the top-level heuristic knowledge of the joint models.Experiments show, even in a cluttered environment with different disturbances, correct results can be acquired by combining top-level model and bottom-level data.