3-D Object Detection Algorithms Based On Dataaugmentation Graph Neural Network

3D object detection algorithm based on the point cloud scanned by lidar is an significant part of sensing algorithm in autonomous driving scene.It is of great significance for autonomous driving in security,and the accuracy of detection technology and will directly determine whether the automatic driving can fall to the ground.With the evolution of deep learning in technosphere and field of market of autonomous driving for several years,many 3D object detection methods have emerged.At present,the most impactive 3D object detection algorithms consist of two categories: Point class and Voxel class.Voxel class uses voxels as the representation of point clouds,which has the longest development history.However,in principle,it will cause the drop of information out of inescapability,and the detection accuracy is close to the upper limit.Point class directly represents point cloud with points,which can retain the original information of point cloud,and graph network method of Point class can combine the regional information of point cloud,which is conducive to the feature learning of unstructured point cloud.This paper first summarizes the seven characteristics of point cloud data,introduces the KITTI dataset,and completes the basic clipping,sampling and other processing work on the point cloud data.Other works are also introduced in detail in this article.For example,we select the vertices through sampling box,and we establish the relationship between vertices and edges within a certain range,and finally combine them into a point cloud graph.The graph is then put into the graph neural network of multilayer perception iterative to update vertex characteristic and predict the category and position through classification loss and location loss.Finally we evaluate the accuracy of detection algorithm through the AP evaluation indexes.Secondly,in order to improve the details of point cloud objects in the detection process,this paper enhanced the local features of point cloud.We summarized a variety of local data enhancement methods used in this article.On the basis of the fact that graph neural network need to select the vertices,a multi-step adaptive point cloud sampling method is proposed based on adaptive block sampling,which can preserve more vertices at a distance in the sparse objects,and suppress the z axis in low or high region where background points take a larger proportion than foreground points.In addition we propose a method of multi-step graph-construction,which leaves much redundancy for accurate vertex selection after the scope of vertices is determined.Then,in order to make use of the global characteristics of the point cloud to make up for the imbalance of the point cloud,this paper uses the global data enhancement method to expand the data volume and prevent overfitting.A global feature enhancement method based on self-attention is proposed,which takes vertex queue as input and makes use of the similarity of objects of the same class to make the output result integrate the interaction features between vertices,so that the features of foreground can be much more obviously distinct from background features and the significance of foreground vertices is improved.This paper also presents a method based on the global properties enhanced visibility classification method,the vertex voxels have a reputation from BEV Angle according to whether it’s blocked or not and result into different categories.A new visibility characteristic is obtained and we fuse the visibility into the characteristics of the vertex queue.Because of the fact that whether the vertex is foreground and whether the vertex of object is blocked are extremely relevant,the ability to identify foreground vertex or background vertex is strongly enhanced.The classification efficiency of the 3D object has promoted greatly.Finally,the ablation learning is performed by using the algorithms we proposed.For the graph neural network,choice of key parameters in multi-step adaptive sampling,altitude area sampling,multi-step graph-construction,self-attention method and visibility method and other aspects,we compare the iterative process of training and testing of the AP values and PR curve,sampling ratio on the test results,and make contrast to the existing algorithm and best graph neural network on KITTI dataset,and at last it is proved that the algorithm put forward by this paper has a great performance in 3D object detection and finally testify the validation of this paper.