Research On Point Cloud Object Detection Method Based On Geometric Features

With the improvement of industrial automation,3D point cloud object detection has been widely used in industrial detection,aerospace,robot navigation,unmanned driving and other fields.At the same time,with the increasing number of spatial data acquisition methods,the acquisition speed of point cloud data is getting faster and faster,and the acquisition method is more and more convenient.Therefore,object detection based on 3D point cloud data has received wide attention from researchers.The purpose of point cloud object detection is to accurately detect whether the object appears in a 3D point cloud scene.This paper comprehensively analyzes the key technologies in point cloud object detection,so as to seek the best pose transformation and improve detection speed.The main research contents of this paper are as follows:(1)Aiming at the current point cloud data lacking standard geometry,we propose a high-quality point cloud data construction scheme for regular solid geometry.In this paper,we use the laboratory Kinect equipment and the EinScan-pro 3D scanner to construct a variety of basic three-dimensional geometric structures including cylinders,cubes,cones and some object point cloud data in daily life scenes.The pre-processing operations such as denoising and repairing the collected point cloud are carried out to obtain the final ideal data set,which is of great value for the future research of the target point cloud composed of geometric primitives.Aiming at the infeasibility of data collection in the complex environment,we propose a method of generating point cloud based on CAD model.Compared with other methods,the point cloud distribution generated by this method is more uniform and the quality of the obtained point set is higher.(2)We focus on the key point extraction and feature description of the object detection algorithm based on point cloud data.In the key point extraction stage,a key point extraction algorithm based on multi-scale and local weight is proposed for the problem of small number of key points extraction and low efficiency of extraction process.It can effectively detect high-quality key points through experimental verification.In the feature description stage,in order to improve the ability of point cloud descriptors,we combine multiple geometric features of local neighborhoods of point clouds.And we carry out extensive experiments on the self-acquired Kinect data and the commonly used 3D model library Retrieval and other data sets,and obtain good results,which verifies the validity and feasibility of the proposed method.(3)We complete the design and development of the point cloud target detection system.Combining theoretical exploration with practical programming,we analyze the feasibility of the point cloud object detection system,and finally realize the functions of point cloud data visualization,key point extraction,and model scene detection result visualization.Then using the model and scene data to demonstrate the overall operational flow of the point cloud object detection system,and verify the effectiveness and practicability of the system.