Robust Surface Abnormality Detection for a Robotic Inspection System

The detection of surface abnormalities on large complex parts represents a significant automation challenge. This is particularly true when surfaces are large (multiple square metres) but abnormalities are small (less than one mm square), and the surfaces of interest are not simple flat planes. One possible solution is to use a robot-mounted laser line scanner, which can acquire fast surface measurements from large complex geometries. The problem with this approach is that the collected data may vary in quality, and this makes it difficult to achieve accurate and reliable inspection. In this paper a strategy for abnormality detection on highly curved Aluminum surfaces, using surface data obtained by a robot-mounted laser scanner, is presented. Using the laser scanner, data is collected from surfaces containing abnormalities, in the form of surface dents or bumps, of approximately one millimeter in diameter. To examine the effect of scan conditions on abnormality detection, two different curved test surfaces are used, and in addition the lateral spacing of laser scans was also varied. These variables were considered because they influence the distribution of points, in the point cloud (PC), that represent an abnormality. The proposed analysis consists of three main steps. First, a pre-processing step consisting of a fine smoothing procedure followed by a global noise analysis is carried out. Second, an abnormality classifier is trained based on a set of predefined surface abnormalities. Third, the trained classifier is used on suspicious areas of the surface in a general unsupervised thresholding step. This step saves computational time as it avoids analyzing every surface data point. Experimental results show that, the proposed technique can successfully find all present abnormalities for both training and test sets with minor false positives and no false negatives.