Multi-sensor fusion for geometric part inspection

Multiple off-the-shelf cameras can be configured to simultaneously provide a large variety of part features that are impossible to capture with a single CCD camera or range scanner. One unsolved problem in using several cameras for passive shape recognition is that of multi-view registration. Registration is the process of associating the feature vectors extracted from the image captured by one camera view with that from another view. This paper describes an unsupervised clustering algorithm used to associate redundant and complementary features extracted from different views of a three-dimensional object for part identification and inspection. The unsupervised learning algorithm ensures that "similar" feature vectors will be assigned to cluster units that lie in close spatial proximity in a three-dimensional feature map. The technique reduces the dimensionality of the input by exploiting hidden redundancies in the training data. During the inspection phase, novel features activate a number of cluster units that have weights similar to the applied training input. If the sum-of-square error between the input and weights of the cluster unit with the strongest response is greater than a predefined tolerance, then the part is rejected. A simulation study is presented to illustrate how the proposed multi-sensor fusion technique can be applied to identifying parts for inspection.