Learning-based method to recognize and localize glassware using laser range images

A system that can be trained to recognize and localize glassware located on a workbench using laser range images is described. The training data consists of laser range images of objects of known classification. The image is preprocessed to isolate a box of pixels corresponding to the object, and then the box is given as an input to a neural network. The range readings from the laser range system deviate significantly from the actual distances to the glassware, and consequently a surface fit to the readings has very little resemblance to the actual glassware surface. Thus, the straightfoward method of fitting a surface to the images and matching it with the surface of known glassware is not feasible. A first version of the system has been developed based on a system of perceptrons, and the tests have been successful on the images taken by a PERCEPTRON P5000 laser range finder.