The Research On Vision-based Surface Detection Of Large-scale Work-pieces And Localization Technique

Monocular vision,as an important branch of visual technology,is becoming increasingly popular in industrial automation,aerospace,medicine and other fields due to its advantages such as simple structure,low cost,high measurement accuracy and fast measurement speed.However,at present,the surface detection and grinding processing of large work-pieces such as locomotives and high-speed trains,are still mainly completed by human labor which is inefficient and the operating environment is harsh.For this reason,this paper studies the surface feature detection and localization of large work-pieces based on monocular vision from three aspects: feature extraction,camera hand-eye calibration and camera pose estimation.Firstly,based on image preprocessing,image segmentation,pixel coordinates of feature areas on the surface of large work-pieces were successfully extracted.For camera hand-eye calibration,the dual quaternion theory was adopted in this paper,and the displacement and rotation in the hand-eye matrix are solved simultaneously by linear least square method,so as to reduce the cumulative error in step-by-step solution.Secondly,after studying the camera pose estimation methods based on epipolar geometry,PnP and plane target,an estimation method applied to large plane was proposed.This method not only has higher applicability,but also can improve the production efficiency.In order to overcome the inherent defects of monocular vision,a novel camera pose estimation algorithm was developed to solve the defect location of large curved surface based on monocular vision.This algorithm which combines ICP and inverse projection based pose estimation,can efficiently solve the 3d coordinates of feature points and camera pose.Finally,the algorithm framework of the large plane and the large curved plane and the physical test are completed respectively in this paper.The overall error is analyzed in detail.The results show that the proposed algorithm can meet the requirements of practical application and improve the level of intelligent production.