The Research Of 3D Object Recognition And Understanding Algorithm In Muti-objective Scenes With Local Surface Features

The ability of robot perception to environment has a significant impact on the robot intelligence,after learning and understanding the environment information,the robot can implement the task planning and decision-making in real time.3D object recognition and understanding in cluttered scenes is a rapidly growing research area.Based on the used types of features,3D object recognition methods can be divided into two broad categories,global feature based methods and local feature based methods.Due to the local feature based methods are more robust to occlusion and clutter in a real-world scene,this paper makes a research on local surface feature based methods of 3D object recognition and understanding in cluttered scenes.A real-time 3D semantic map generation method based on the open source Point Cloud Library(PCL)is presented in this paper.The main contents are as follows:(1)Building a point cloud model database for modular manipulator sorting platform operation's objects.Removing the noise point and outliers in the point cloud data through the filter preprocessing,designing an improved region growing segmentation algorithm to achieve the extraction of object clustering,using Intrinsic Shape Signatures(ISS)algorithm to detect clustering's key points and calculating the corresponding local feature descriptor Rotational Projection Statistics(RoPS),establishing the k-d tree sequence table.With these methods,this paper builds the model databases.(2)In order to improve the efficiency of modular manipulator sorting operation,this paper proposes a real-time 3D semantic map generation method.This method contains a lot of phase algorithms,such as filter preprocess,3D key point detection,local surface feature description,feature matching and so on.After analyzing the merits of each algorithm,one complete 3D object recognition and understanding algorithm is proposed.The experiment tests the method whether satify the requirements of real-time acquisition and high accuracy.(3)The modular manipulator sorting system hardware and software architecture is constructed.The real-time 3D map generation algorithm,human-robot interaction reasoning algorithm and language understanding and manipulator control algorithm are encapsulated into software modules,and the whole system is tested intensively.The experiment results verify the effectiveness of the above scene recognition and understanding algorithm.Finally,this paper does a conclusion of the research work,analyzes the innovation points and research value,and proposes the further research suggestions.