Object 3D Modeling And Pose Estimation Based On Kinect

In order to satisfy the requirement of the service robots’ grasping and loosing common object in the unstructured indoor environment, object 3D modeling and pose estimation of object in the current state is a necessary. To solve this grasping problem, the paper proposes a method to rebuild 3D models of common objects, and build a strategy for object recognition and localization based on the local invariant geometric feature available from the Kinect sensor. This article carry out research work as follows:Firstly, in order to obtain object model with high degree of accuracy, the analysis of depth accuracy and noise with multi-view raw RGB-D cloud provide optimal sampling area. On basis of noise experiments, mechanism of point cloud preprocessing is proposed to eliminate depth noise and color noise.Considering the robot observation angle problem, Sampling multiple view of object as a foundation for subsequent object feature template construction, and establish the 3D model. Based on the texture information and geometry information, using the SIFT matching and singular value decomposition method for initial registration in terms of disorderly and partially overlap of adjacent views, the article improve ICP algorithm to realize accurate registration. Finally, closed-loop optimization strategy is proposed to greatly reduce error accumulation when obtaining accurate rotation and translation motion relationship among multi views and constructing the complete object model.Object recognition and pose estimation. Acquiring the multi view motion relationship of rotation and translation in the process of object 3D reconstruction, a sparse expression of object model based on local invariant geometric feature is proposed. The method of Euclidean cluster extraction is selected to divide an unorganized cloud of clutter scene, template matching with object’s template database can realize object recognition and pose estimation. Lastly, according to the request of robot grasping and object recognition and position calculation results, the location and pose estimation of object model coordinate system relative to the robot’s coordinate system is calculated.The above strategy is validated through the experiment, this paper reconstruct the 3D model of actual object. Improved ICP algorithm and the closed-loop optimization strategy are proposed,which providing technical support of ensuring accuracy of location and pose information to meet the requirements of robot grasping.