Research On 3D Object Detection Technology Of Laser Point Cloud Based On Convolution Neural Network

With the continuous progress of data acquisition and storage technology,the significant improvement of computer data processing ability and computing level,and the rapid development of artificial intelligence and automobile industry,automatic driving technology has become the focus of the industry.Object detection,as a basic component of autopilot perception,has also become a research hotspot.Compared with 2D object detection,the position estimation of 3D spatial object has more important meaning in practical application.By estimating the 3D position of objects in space,intelligent vehicles or robots can accurately predict and plan their behaviors and paths to avoid collisions and violations.The point clouds captured by Li DAR has many advantages,such as high accuracy of depth information and good consistency of 3D spatial scale and so on.Therefore,point clouds has become an important data form of 3D object detection.However,3D object detection based on laser point cloud faces more challenges,mainly reflected in the need for accurate estimate the 3D position and size of the object.In this thesis,the convolutional neural network is applied to Li DAR point cloud 3D object detection,and related theoretical and technical research is carried out in depth,and the following results have been achieved.First of all,for 3D object detection of pure point cloud data,a voxel feature learning algorithm based on point cloud voxel representation is proposed.The proposed algorithm makes full use of the characteristics of voxel interior points.And based on the cloud features of adjacent voxels,the algorithm uses trilinear interpolation algorithm to obtain the same neighborhood features as the number of interior points of voxels.Finally,the deep fusion of point features and neighborhood features is realized by feature splicing,full connection layer and pooling layer.The voxel feature learning network extracts robust voxel features with both point feature and neighborhood feature through end-to-end learning.Secondly,a multi-scale feature extraction and fusion network is proposed to detect 3D objects with BEV feature map.In this network,firstly,the down-sampling is used to extract feature maps with different scales from the BEV map obtained by voxel feature learning network.Then,the size of feature maps is unified by using the upsampling.And these feature maps are spliced together to complete multi-scale feature fusion.The fusion feature effectively improves the detection accuracy,for it contains both the details information of the low-level feature map and semantic features of the high-level feature map.Finally,in view of the inconsistency between the design of loss function in 3D object detection and the the evaluation standard of detection accuracy,this thesis analyzes and studies the problem.As a result,a 3D generalized intersection over union loss function is proposed,which is combined with other losse functions to complete the optimization task of the model.By using this loss function,the direction accuracy of the prediction frame can be improved,so as to improve the accuracy of object detection.