Research On Key Technologies And Comprehensive Application Of Machine Vision Pose Measurement

As a sub-domain of machine vision,pose measurement is widely employed for production line automation,stereo scene reconstruction,non-contact survey,pilotless driving and intelligent logistics.From the perspective of accuracy,robustness and adaptability of pose estimation system applied in industrial fields such as four-wheel alignment and unmanned stockyard,relevant theories and key technologies of pose measurement via multi-camera and laser scaner are researched as follows:A noval approach for localizing sub-pixel corners is proposed,by establishing an ideally continuous chessboard corner model,as a function of corner coordinates,rotation and shear angles,gain and offset of grayscale,and blurring strength.The ideal model is evaluated by a low-cost and high-similarity approximation for sub-pixel localization,and by performing a nonlinear fit to input image.Under stable and variable illumination conditions,the proposed method demonstrates a betther performance than the other representative algorithms for sub-pixel localization.A self-checking technique is also proposed by investigating qualities of the model fits,for ensuring the reliability of addressing Pn P problem.The outlier recognition and elimination are independent of Pn P optimization process.When the target features are polluted by light,the technique can quickly and accurately identify and eliminate outliers,improve the reliability of camera pose estimation and ensure real-time performance.A point-cloud registration method for large scene is proposed based on plane feature extraction.On the premise of ensuring the matching accuracy,it focuses on the matching efficiency of the algorithm,which has certain advantages when applied to the registration of large scene point-clouds.A global calibration method for non-overlapping multi-camera system is proposed,by taking into account the non-uniformity of the calibration device,and optimizing the position and attitude by calculating the difference of the position and attitude of each calibration postures with respect to the others.The results of simulation and practical experiments show that the global parameter calibration accuracy of the quadratic optimization can be improved significantly.On this basis,the global calibration problem of arm-borne vision positioning system is analyzed,and the method of solving calibration parameters between joint arm and vision system is discussed.The methodologies of multi-camera pose measurement are integrated and applied in machine vision-based four-wheel alignment system.For addressing the problem that four-wheel alignment is sensitive to validities of steering wheel reset and target installation,an accurate baseline estimation method is proposed,by elaborating the principle and solving process,in conjunction with geometric constraints of four-wheel motion association.The proposed method is verified using synthetic data,and implemented in a machine vision-based four-wheel aligner with conventional hardware organization;experimental results show that alignment parameters can be measured in a higher accuracy.The methodologies of pose measurement via point-cloud scanning are integrated and applied into unmanned operation and pile localization system for enclosed stockyard.Aiming at the problem that the existing calibration methods for open stockyard are no longer applicable in the fully enclosed owner site,a global calibration model is established based on the angle and displacement sensor network of the equipment.On-site test results show that the accuracy of unifying point-clouds obtained using three-dimensional laser scanner can meet the requirements of unattended operation.