Research On Texture-less Objects Recognition And Pose Estimation Based On Kinect V2 Sensor

With the implementation of the manufacturing power strategy "Made in China 2025",the vision-based robot technology has been widely used in the fields of object handling,assembly and so on.However,the intelligence of object detection and pose estimation is not good in practical production,especially in the unstructured environment,the background confusion,mutual occlusion and other scenarios are still a chal lenge to the existing algorithms.The 3D detection system with high performance vision sensor can obtain geometric information of the target workpiece,but the expensive price restricts its application in practical production.In this paper,We use the low-cost Kinect V2 sensor to obtain color images and depth images.We research on object detection and pose estimation in unstructured environment.The contents and innovations of this paper are as follows:1)Research on getting accurate object point cloud based on Kinect V2 sensor.The planar object calibration technology is used to establish the transformation relationship between the depth image and the 3D point cloud.There are defects such as discrete noise and missing holes in depth image acquisition based on Kinect V2.The edge features of the original image are protected by guided filter while removing noise,and the effects of adjusting parameters and window radius on image quality and boundary contour are analyzed,and the optimal selection parameters are proposed.Based on RGB image,the missing holes of deep image is processed by joint bilateral filtering,influence of the parameters is analyzed,and the number and quality of the holes filling are tested by experiments.2)Research on template matching algorithm based on multi-featured 3D objects In view of the shortcomings of existing Line MOD multi-feature template matching algorithms which block the foreground of the target,can not recognize and detect the symmetry of the geometric structure of the target,it has the disadvantage of repeated detection results.The methods of non-maximum suppression and hysteresis threshold are used to refine the edge contour and obtain the optimal contour of the target object and the current scene.The number of improved template features is larger than the feature of LineMOD template matching calculation.and the computation time will be increased.Therefore,the image pyramid search method is used for fast template matching.Then,the similarity function of the template matching is redefined for the improved method,and the degree of acquaintance between the calculated template image and the neighborhood window of the scene image is detected.3)Improved Iterative Closest Point(ICP)accurate registration algorithm.Fro m the practical application,on the basis of the classical ICP algorithm,we improved from the aspect of how to improve the registration speed.The voxel grid downsampling method is used to downsample the point cloud data.Then use KD-Tree to establish the topological relationship for the downsampled point cloud data,so as to quickly find the corresponding point pair.Finally,the Euclidean distance threshold is used to eliminate the error point pairs in the corresponding point pairs of the matched templat e point cloud and the object point cloud,so as to reduce the registration error of the point cloud.