Precision Positioning Of Workpiece Target Using Single Vision In Complex Scenes

Machine vision has played an increasingly important role in industrial applications with advances in technology. The targets should be accurately located,when the industrial robot sorts workpieces.Registration is one of the main means of solving industrial targets accurate recognition, and image registration can be used to identify the target in sub-pixel accuracy. Image registration is to find the position of an image in the other image. This paper studies image registration problem in the industrial field. Because the registration problems could not get the closed-form solution in a complex environment, in the complex industrial scene, the false matching rate of target recognition is often very large.How to improve the correct rate and accuracy of the registration in complex scenes is one of the key problems that need to be resolved at this stage.Firstly, we used orientation chamfer matching based on twin-threshold segmentation to roughly recognize the target to find the crude match point. And then we used registration algorithm in this paper to accurately recognize the target in the complex environment. Finally, the experiment results showed the performance of the three improved registration algorithm in the complex scene using coordinate transformation.The main contents and work as follows:(1) We developed the algorithm of orientation chamfer matching(OCM) based on twin-threshold segmentation. Firstly first segmentation was carried out to the target image under the complex background using Otsu method. Secondly second segmentation was carried out to the target image using iterative threshold method with the purpose of segmenting the targets form the complex background. Finally OCM was used to recognize the targets. the work-piece`s average recognition rate reached 72.75% and the right match rate of this algorithm is 99.25%.(2) This paper focuses on three traditional registration algorithm, respectively named SIFT feature image registration, Gauss-Newton image registration and iterative closest point(ICP) image registration. Then we developed this three algorithm and got three new algorithm, respectively named constrained SIFT feature image registration, trust region damped Gauss-Newton image registration and similar SIFT denoising truncated iterative closest point(TICP).Firstly, according to the specific characteristics of the work objectives we made some improvements for SIFT operator, adding a constraint SIFT feature alignment concept. We used the high robust Shape Match template matching to find the rough geometric transformation information, and then find the matching feature point pair in the template image and the image to be registered. We used the rough geometric transformation information to eliminate the error matching point pair,and then we estimated geometric transformation parameters using random sample consensus?Simulation results show that the proposed algorithm the mean error in x and y direction respectively were 0.030 pixels and 0.041 pixels, and the mean error of rotation angle is 0.11 degrees. The right match rate of this algorithm is 90.7%.Secondly, this paper studied some shortcomings of the Gauss-Newton method: if the initial point is far from the target point, the Gauss-Newton method could not find the extreme value because of the too large step size and Gauss-Newton method is easy to fall into local solution in complicated scenarios. In this paper we developed Gauss-Newton method and we got the trust region damped Gauss-Newton method. In the course of each iteration we set confidence interval, then we used damped Gauss-Newton to find the extreme value. Simulation results show that the proposed algorithm the mean error in x and y direction respectively were 0.0345 pixels and 0.0736 pixels, and the mean error of rotation angle is 0.0984 degrees. The right match rate of this algorithm is 86.7%.Thirdly, we develop ICP aimed at the problem that ICP could fall into local solution in a complex scene. We got the similar SIFT denoising truncated iterative closest point method. We used similar SIFT operator to remove the noise point. Then we used TICP to register the image. Simulation results show that the proposed algorithm the mean error in x and y direction respectively were 0.043 pixels and 0.0542 pixels, and the mean error of rotation angle is 0.0984 degrees. The right match rate of this algorithm is 90.0%.(3) We designed the target real position determination experiment in complex scenarios to measure the performance of the proposed algorithm in this paper. Aim at the workpieces of three different categories, real experimental results showed that: the x direction average error of constraints SIFT feature matching algorithm,trust region damped Gauss-Newton image registration and similar SIFT denoising truncated iterative closest point method, respectively, is less than 0.6mm, 0.4mm and 0.5mm. And the y direction average error, respectively, is less than 0.7mm, 0.55 mm and 0.5mm. And the running time is respectively 0.18s~0.22 s, 0.88s~0.94 s and 0.50s~0.57 s. So this three algorithms can be applied to the workpiece accurately identify in complex scenes.