Research On Vision Localization Based On Irregular Object

During the Modern manufacturing, in order to achieve processing the objectsautomatically, Machine must obtain the position and orientation of objects in theworkbench in advance. Only with this information, the machine can know where theobject is and which orientation it should process. In this paper, the Applicationbackground is the irregularly shaped leather embossed production. In order toachieve embossing the irregularly shaped leather automatically which are placed onthe working platform at random. In this paper, with the help of a monocular visionsystem, it extracts the position and orientation of irregular objects by the means ofvisual positioning. In addition, as for different types of objects in the actualmanufacturing, it may be require for different operations. For example differentshaped leather may be embossed on different types. So, we also need to distinguishtheir categories. For the above problems, we did the following work:Firstly, this paper extracts the corner of the target image. Because accuracy ofthe position is not very high, we need to find the sub-pixel position. There areseveral calculated points that are not projection of the real points of the target. Weselect the right points with the way detecting the degree of symmetry. By the way ofsorting the corners, we set up the corresponding relations between the corners’coordinates in the image and the corners’ coordinates in the real target. Thenaccording to the camera model, we calibrate the camera. According to the target andtable coordinate conversion relations, we eventually establish correspondingrelations between image plane and the table plane.Secondly, in this paper, we chose the appropriate pre-processing method topreprocess the image, eliminating the noise and the segment objects from thebackground. Then with the help of the geometric moments, we achieve to extractthe position and orientation of the irregular objects. But it will take a lot time. Thispaper adopted a combination of connected component labeling and interlacedseparated Column, and it realized to identify multiple objects and reduce theprocedure time a lot.Thirdly, this paper analyzes the feature extraction method for the identificationof irregular objects, and selected Hu invariants moment as the irregular objects features. We adopted the variable step BP neural network with adaptive momentumfactor to classify and achieved recognizing different irregular objects.