Methods On 3D Object Detection Of 3d Point Clouds And Its Application In Industry

Object detection is a hot topic in the field of computer vision.Due to the lack of depth information of visual sensors,it is difficult to detect the three-dimensional shape of objects for traditional two-dimensional detection algorithms.In contrast,using the point cloud depth information provided by the RGB-D sensor,the 3D object detection algorithms can output the distance and shape information of the object.In this way,the environment can be perceived in three-dimensional space to meet the needs of higherlevel target detection tasks in the fields of robot vision and augmented reality(VR).The point cloud-based object detection algorithms utilize the geometric information of the original data,which have significantly improved the object detection performance.Among them,the Vote Net detection network,which uses the hough voting method with the original point cloud as input,has achieved significant accuracy improvement on the public data set.However,it is found that in the Vote Net algorithm,point clouds from walls,ground and adjacent objects interfere with the detection of "targets",which seriously affects the detection performance of the network.Based on the above problem,this thesis proposes an optimized solution.The experiments show that the work in this thesis can effectively improve the detection performance.At present,the training and testing of 3D object detection algorithms are limited by common indoor environments in public datasets,and cannot verify the detection performance of uncommon objects in industrial scenarios.Therefore,this thesis carried out experiments in several specific application scenarios to verify the feasibility of Vote Net in practical applications.Finally,take the auto repair shop scene as an example to carry out the target detection experiment in the industrial scene.In addition,we designed a complete processing flow from data collection to data set training,which can provide a reference solution for the application of practical scenarios.The main work of this thesis is as follows:(1)A simple and effective optimization scheme is proposed for the problem of the influence of interference points in the point cloud data in Vote Net.In the optimization scheme,which adds a filter module to the proposal module,we use a pre-generated simple bounding box to filter the point set.It is improved that the optimization scheme can effectively remove the interference points,so as to improve the detection performance of the algorithm on the public data set SUN RGB-D.(2)In view of the difficulty in acquiring point cloud data and the poor quality,in this thesis a point cloud data acquisition platform was built,which is equipped with a RGB-D camera,a two-dimensional lidar and the robot platform.It can be remotely controlled through a mobile phone,which supports convenient and fast point cloud data acquisition.(3)In view of the problem of limited public datasets,multiple datasets were constructed in this thesis,and the application experiments of Vote Net detection network from common scenes to industrial scenes were carried out.This thesis provides a complete point cloud data set construction process,which is processed with reference to the public data set SUN RGB-D.And in the data labeling work,a technical solution for fast dataset labeling is proposed.In the Vote Net detection network application experiments in this thesis,a series of datasets are constructed for training.First,experiments are carried out in conference room scenes,office scenes and laboratory scenes to analyze problems and improve the data processing process.Finally,we take the auto repair shop as an example to carry out 3D object detection applications in industrial scenes,so that the feasibility of the network application in multiple scenarios is verified.