Research And Implementation Of Key Technologies In Bidirectional Vision-guided AGV

Automated Guided Vehicle（AGV）has been widely used as transport for materials in modernmanufacturing system. Theoretically, vision-guided AGV has a good application prospect, but it hasnot been widely equipped as electromagnetic guided and laser guided AGV. The key issues are thatthe real time ability, robustness and accuracy of the vision-guided technologies have to be improved.This paper reviews the current researches on AGV and vision-guided technologies, and presents a setof new methods mainly to solve the following four key problems for bidirectional vision-guided AGV.System error of vision-guided AGV comes from two aspects, the image distortion and the poseerror of camera compared to the motion control coordinate system, which can be effectivelyeliminated by accurate system calibration. Considering the system structure characteristics ofbidirectional vision-based AGV, a system calibration technology based on static state and motion stateis proposed. Firstly, the internal parameters, radial distortion parameters and external parameters ofcamera are estimated by using planar patterns in the static scene. Then a union model for correctingdistortion is built for three individual image distortion models. Finally, the rotation and translationparameters between the corrected image coordinate system and AGV motion control coordinatesystem can be calibrated in two motion cases, so the camera pose is estimated accurately in AGVmotion control coordinate system. Experimental results show that the technology has the features ofgood flexibility and high accuracy.Vision information processing must be built on the basis of the authentic and accurate featureextraction for guide path. In practice, the problems of non uniform illumination and high lightphenomena caused by the coaxial annular LED array light used for the far scene and large field mustbe solved. This paper builds an illumination model for the annular LED light at the far scene. Themodel parameters are estimated by using the nonlinear least squares Levenberg-Marquardt algorithmfor the monochromatic diffuse reflection template image. Another method based on irradiance averagenormalization is proposed to remove the influence of non uniform illumination. In order to solve thehigh light phenomenon, the blue color chrominance Cb of YCbCr image is complemented, and thenthe bilateral filtering algorithm is used for image enhancing, which improves the robustness of featureextraction for blue guide path.After image distortions correcting and non uniform illumination image enhancement, the centerlineof bidirectional guide-path is extracted by using the color image pre-processing algorithm. A newmethod based on statistical characteristics of curvature estimation is presented to adaptively classify the models of straight line, arc turning and other non-ideal paths in this paper. Firstly, the motioncharacteristics of bidirectional AGV and characteristics of three types of planar curve are analyzed.Then in terms of the target accuracy of measurement, a classification method based on the curvatureestimation is proposed to classify path models into straight line, arc turning and non-circular turning.Finally, the curvature estimation based adaptive weight fitting method is proposed for the parametersregression of the three models. The refined parameters are used to compensate system errors, whichimproves the accuracy of vision measurement.Guide-path for vision-guided AGV can be divided into bidirectional path and cross path. A newhierarchical recognition method based on the real-time ability of knowledge acquisition and thesimilarity of classes is presented to robustly recognize the cross path models for bidirectionalvision-guided AGV in real time by the combination of the rough set theory and the multi-class supportvector machine The knowledge granularities conception and the hierarchical reduction rules of roughset theory are both used to obtain the minimum decision rule, which effectively reduce the complexityof the classification. To improve the robustness of recognition, the learning method of safe area forclassification is presented, which makes the linear inseparable uncertain problem become linearseparable conditionally. Finally, the tests and experiments at various environments verify the validityand reliability of the method.A prototype of bidirectional vision-based AGV, NHV-II, is implemented based on the study of thetheory. TMS320DM642DSP is used as the vision system processor of NHV-II. The multitaskalgorithms related to video capturing, video processing, and communications, etc are performed onthe real-time embedded system DSP/BIOS. The RFID reader, the industrial wireless local areanetwork communication and the center control workstation are integrated to implement map building,station identification, path planning, scheduling and state monitoring of the AGV system. Finally, themethods and technologies presented in this paper are tested by using NHV-II on different environmentconditions. Experimental results show that the problems of real time ability, robustness and accuracyof the vision-guided technology are improved significantly, which lay fundaments for the promotionof vision-guided AGV.Finally, the main research results of the thesis are summarized. The difficult problems andvaluable directions for further researching on vision-based AGV are pointed out, which are expectedto be solved and improved gradually in future studies.