Research And Application Of 3D Object Detection Technology Based On Outdoor Point Clou

With the rapid development of 3D deep learning,3D object detection algorithms based on point clouds have received great attention,and there is nothing more important than improving the accuracy and real-time performance of 3D object detection algorithms,so how to make full use of point cloud data for accurate and fast target detection has received widespread attention.In this paper,3D object detection techniques based on outdoor point cloud are investigated,and the specific research work is as follows.(1)A combined clustering and classification method for 3D object detection is designed.Firstly,the normal vector of each point in the point cloud is calculated,and the non-ground points are initially filtered by the normal vector.Then,the RANSAC method is used to fit the plane of the initially filtered point cloud to obtain the plane where the ground points are located,so as to filter the ground points,reduce the computational effort of the subsequent clustering algorithm and improve the clustering accuracy.The DBSCAN algorithm is then used to cluster the point cloud after the ground points have been removed,and the 3D enclosing box information of each cluster is calculated by axis alignment.The clusters are then used as input to the improved Point Transformer model to obtain the categories of the different clusters and complete the final detection.In the improved Point Transformer module,the features of each point in the point cloud are extracted using a graphical convolutional network instead of MLP,in order to solve the problem of lacking feature information of neighbouring points after the points in the point cloud have been up-dimensioned by the MLP layer.The experimental results show that the method has better point cloud classification accuracy and generalisation capability for 3D object detection.(2)A detection method that combines adaptive point features with voxel features is proposed so that the network model can both make effective use of each point feature and obtain a large perceptual field.The network is divided into two stages.In the first stage,the rasterised point cloud is sparsely convolved in three dimensions several times and the voxels obtained from the last layer of sparse convolution are projected to obtain a bird view,which is used to generate the 3D suggestion box.Considering the random nature of the key points obtained by the farthest distance sampling,this paper adopts a deformable convolution scheme in the second stage to adaptively adjust the position of the key points obtained by the farthest distance sampling,so as to make the features of the key points more distinguishable and representative,and at the same time,to address the problem of mixing the target and background points in the sampled key points,the Self-Attention module is proposed to highlight the features of the former points and suppress the features of the background points.The method also proposes a Self-Attention module to highlight the features of the former points and suppress the features of the background points.The experimental results show that the method can effectively improve the accuracy of 3D object detection.(3)A 3D object detection system based on outdoor point clouds is designed and implemented.The task analysis and process modelling of the data acquisition module,data pre-processing module and 3D object detection module are carried out,and finally the effectiveness of the 3D object detection method in this paper is verified in an outdoor point cloud environment.