Research On The Technology Of Object Reconstruction And Target Recognition And Orientation Based On Kinect

Since the aggravation of aging society is becoming increasingly serious,the family service robot has gained a a broader application prospect.Grasping of common objects in complex environent is an indispensable function of service robots,of which object recognition and location technology is a prerequisite.Addtionally,a complete 3D object model is an important prior knowledge in the pose calculation process.Therefore,the aim of this study is to design an user-friendly object reconstructing method and an recognition and location scheme to complex environment which includes the construction of visual perception system,the process of data preprocessing,the reconstruction of object model,and the algorithm of object recognition and location in the indoor scene.First of all,apply point cloud data as the carrier of environmental perception and building Kinect point cloud data acquisition and processing system.Then construct the measuring coordinate system and the linear and nonlinear models of the camera based on the imaging process of the camera.Zhang's calibration method is used to solve the internal parameter matrix,distortion parameter and outer parameter matrix of color camera and depth camera.At last,the Kinect camera calibration is completed,and the scene color spot cloud is collected in real time.Secondly,according to the application requirement of accurately extracting the point cloud segments from the environment,study the preprocessing algorithm of point cloud.By analyzing the types of noise generated during Kinect point cloud shooting,a multiple filtering algorithms are proposed.Use the statistical analysis method to remove the edge noise of point clouds,the principal component analysis method to extract the surface normal and curvature of the point cloud,and the improved bilateral filtering algorithm to smooth the small noise on the surface of the point cloud.Then the random sampling consensus algorithm is used to remove the unsupported plane,and the target fragments are extracted by smoothness Euclidean distance clustering in order to obtain the raw materials of 3D reconstruction.The purpose of separating the target from the background is finally achieved.Thirdly,the 3D reconstruction technology of the object is studied.The traditional RANSAC pre stitching algorithm is improved by using the SIFT3 D algorithm to extract the key points of the point cloud,and the local FPFH feature matching to avoid repeated calculation in the pre matching process.In this way,the robustness of the stitching algorithm is enhanced.Precision registration algorithm based on neighborhood features is used to accurately stitching,and the probability of mismatching is effectively reduced by curvature analysis,and the matching of local minimum or error results is prevented.Experimental results show that the algorithm in this paper has higher accuracy and reliability compared with the traditional ICP stitching algorithm,and has fewer iterations.In this chapter,the complete 3D structure of objects to be detected is obtained through two frame registration.Finally,the object recognition and location algorithm is studied.In order to avoid the high complexity of KNN algorithm and the influence of irrelevant dimension on classification accuracy,the method of pre clustering partition and computing the weights of each dimension is introduced to reduce computation and increase the accuracy of recognition.Then according to the different advantages of local features and global features,a recognition and location algorithm for feature fusion is proposed.In the off-line phase,the object coordinate system is constructed,and the multi view CVFH global features and SHOT local features are extracted.In the online phase,the global feature is used to identify the types of point cloud segments.Only the local feature of the effective point cloud fragment will be extracted,by which invalid calculation reduced.By splicing point clouds and models with coordinate system,the six degree of freedom position matrix could be calculationed to complete orientation proccess.