Research And Implementation Of Parts' Classification And Recognition System Based On Multi-view Stereo Vision

In modern industrial manufacturing,with the advancement of image processing and computer technology,some work done by human eyes is being gradually replaced by machine vision technology.At present,in the visual sorting application of parts,it is still based on 2D image to classify and position the parts,which is sensitive to the change of the height and inclination of the parts,and the industrial production process is still not flexible enough.To solve the above deficiencies,this paper develops a set of parts classification and recognition system based on the multi-view stereo vision.Firstly,from the perspective of multiview stereo vision,the hardware environment of the system as well as the overall software architecture is determined.Secondly,aiming at the situation that the binocular camera can not cover the whole field of vision of the target parts,a trinocular stereo vision technology based on binocular disparity method is used to do the 3D reconstruction of parts.Then in terms of parts' classification and identification,the paper uses both surface-based matching method and Support Vector Machines method to do the research on classification and identification of parts' 3D models.After that,the software design of the system,which includes the image acquisition module,camera calibration module,3D reconstruction module,surface-based matching module and Support Vector Machines module,is designed and implemented.Finally,the results of 3D reconstruction,parts' classification and recognition and pose location in the designed system are analyzed by experiments.The 3D reconstruction experiments analyzed twenty 3D reconstruction models of two different size parts.The results show that the precision of 3D reconstruction can reach about 1mm.By performing one thousand times experiments of the two kinds of classification and recognition methods in the case whether or not the parts are overlapped with each other or there are abnormal parts,the results show that the recognition rate of the two methods both can reached above 95%,and two methods have different applicability under different circumstances.In the positioning experiment,by comparing the positioning result with the actual position of the parts,the validity of the calculation method and the accuracy of the calculation result are verified.