Research On 3-D Object Detection Algorithms Based On Point Cloud Feature Enhancement

The task of autonomous driving has received much deserved attention and faces many challenges in technology.It is very important and difficult to understanding the 3D environment effectively in the field of autonomous driving.There are two main types of data collected by autonomous vehicles during driving,the RGB image data and the point cloud data.The RGB image data can be effectively processed by the convolutional neural network,but how to process the point cloud data still remains an open problem.Otherwise,the difficulty of detecting different-scale objects exists both in RGB data and point cloud data,there is no effective method to solve the multi-scale problem.This article use novel graph network algorithm to design the network structure and loss function to solve the multi-scale problem in scenes of autonomous driving,and achieve the target detection of point cloud data.Firstly,the KITTI dataset is introduced and the node graph is built in the point cloud data.In order to reduce the Computational complexity,the point cloud is downsampled to build a simpler graph,and the key parameters of the graph construction are discussed.Then,a multi-layer perceptron MLP is used to build a GNN network to process the point cloud graph.The point cloud graph is generated by using the point cloud as a vertex and the relationship between the two point clouds as an edge.After multiple iterations of the same network,the features are obtained.The classification head and the location-regression head use Focal loss and Huber loss as classification loss and position regression loss,respectively.Secondly,in order to solve the problem of point cloud sparseness in distant areas caused by fixed-size voxel grid,a node graph generation method for multiscale voxel sampling grids is designed.The sampling voxel is gradually reduced according to the distance and the number of point clouds in sparse areas can effectively increase.Aiming at the problem of the increase in the number of point cloud nodes brought by the above method,a regional suppression scheme is proposed to reduce the number of sampling points,so as to achieve the balance of the number of sampling point clouds and the detection results.Then the visualization of the proposed method is visualized and compared with the original method of fixed voxel sampling grid.Then,for the multi-scale problem of point cloud,an updateable point cloud map combined with self-feature alignment loss function is designed.An updateable point cloud map is designed based on the distance of features on the feature space.Due to the difficulty of searching for neighboring points in the feature space,an unsupervised loss function with self-feature alignment is applied to reduce the difficulty of search.Finally,the curves of the average precision and the recall rate during the network convergence process are shown.Finally,the ablation study is performed on the proposed novel algorithms,which was divided into three parts: key parameters of graph network,node graph generation method of multi-scale voxel grid sampling,self-feature alignment loss function and updateable point cloud graph.The average precision and the recall rate of the training process,as well as the experimental results in the valid dataset,fully prove the effectiveness of the proposed novel method.And we compared our method with the state of the art algorithms in the published KITTI dataset,the most advanced results are obtained.