Research On Kinect-based 3D Pattern Recognition Technology

The ability of the robot to perceive the external environment determines its intelligentialize,and machine vision is main way that robot get information from outside world,so it has always been one of the most active scientific research filed.With the rapid development of 3D camera,and the arise of open source Point Cloud Library in recent years,based on point cloud image of 3D object recognition technology become a new research hotspot in machine vision.The basic steps of 3D object recognition based on point cloud include point cloud data acquisition,point cloud data preprocessing,point cloud segmentation of object,point cloud feature extraction,build feature model database,match and recognition etc.this paper detailed analysis and studied theory and implementation of each step,while adding two steps include recognition results verification and real-time display results,propose a Kinect-based 3D objects real time recognition method,it can recognize individual object type and pose,cluster objects match and recognition.The main research contents include:Point cloud data acquisition use Kinect through the Kinect SDK and IO module of PCL capture point cloud data in real-time,then use point cloud preprocessing steps like Pass Through filter,Voxel Grid filter,Remove outlier points reduce subsequent processing of the data.Object point cloud segmentation through the analysis of the different cluster segmentation methods,propose a combination segmentation method which based on random sampling consistency and Euclidean distance,then acquire data in real-time segmentation,experiment results show that the combined segmentation method is very efficient.Point cloud feature extraction through the research on 3D histogram feature descriptor extraction method and experimental analysis,extract object feature to recognition,and match the corresponding point cloud model of object to build the feature model database.Match and recognition through the analysis of based on 3D descriptor correspondence grouping recognition method,propose a recognition method that use local descriptor retrieval feature model database,globe descriptor match correspondences.Recognition verification use align registration and hypothesis verification automatic evaluate recognition results and real-time visualization.The real-time recognition experiment include single object recognition and cluster objects recognition.The first experiment in two states of Kinect stationary and moving respectively calculate the real-time recognition rates of individual objects.The real-time recognition experiment of cluster object includes the object's non-occluded recognition and occlusion recognition.Eventually through experimental results and evaluation criterion estimate if the method in this paper meet the real-time recognition of objects requirement in the room,and whether they can locate the target object,lay a foundation for future roboti precise grasping objects,Obstacle avoidance,navigation.