Abstract: The active visual sensor system based on structured-light is an effective method to detect weld seam profiles used in robotic automatic welding with thick steel plates. A general algorithm is presented to extract the typical laser stripes via visual features from the images with or without entire arc regions, which contains three steps as follows. Gabor filtering with appropriate filtering angles is used to produce the orientation feature map, which is segmented by Otsu algorithm and clustered via nearest neighbor clustering. The clusters whose lengths are over the average one are kept whereas those clusters with small lengths are eliminated, through which the approximate major parts of the laser stripe are first extracted. The first part of the laser stripe is discerned through the visual features of the laser stripe from the above approximate extraction result. By using the discerned part a competitive strategy that can guide how to discern subsequent parts of the laser stripe step by step, is developed. The strategy depends on two factors-the Euclidean distance between the candidate clusters and the last discerned cluster, and the lengths of every cluster, through which the major parts of the laser stripe can be then extracted. A rule of how to identify the small parts that lie between two adjacent major parts of the laser stripe is investigated. To verify the effectiveness of the proposed method, two types of images that are captured in the welding process with T-joints and butt joints, are applied to the tests. The proposed method affords the possibility of integrating different functions into an intelligent robotic welding system for weld profile extraction with different joints.

Key words: laser stripe extraction, visual features, thick-steel-plate welding, clustering, orientation feature map