Manipulable Object Recognition Research Based On 3D Point Cloud

During the patrol inspection of nuclear power plants, because of the unbearable hazards of nuclear radiation to human bodies, it’s necessary to substitute a robot for a person. A robot has effective vision capabilities as a precondition for its realiable operations.This article studies the valve detection and recognition methods based on Point Cloud Library (PCL), builds the point cloud data model of valves for nuclear power plants, provides a detailed description of the valve point cloud detection process, such as point cloud pre-processing, segmentation, feature extraction and matching, and further recognizes the type of the target valve after detecting the valve’s point cloud data from the test scene. The main contents of the paper are as follows:1.Build Kinect point cloud acquisition interface using MFC (Microsoft Foundation Class Library), describe process of filters, segmentation and features involved in point cloud pre-processing stage in detail, review the advantages and disadvantages of each algorithm, and then propose an improved European clustering segmentation algorithm for point cloud image segmentation.2.Introduce and compare the local descriptors and global descriptors for point cloud, ultimately use the Viewpoint Feature Histogram (VFH) to capture the point cloud features and build three kinds of valve point cloud model.3.By respectively compared with the test scene point cloud’s cluster features and the model database’s features, we detect the valve point cloud. Then, building the SVM(Support Vector Machine) to recognize the valve type and using the cross-validation method to optimize the model to achieve a higher classification accuracy. The detection part and recognition part work together to complete the process from detecting the valve in the test scene to recognize the valve type.