Research On 3D Point Clouds Density And Application In 3D Object Detection Based Deep Learning

Object detection is an important branch in computer vision,which aims to perceive the surrounding environment through machines,including the classification and positioning of objects.Objects,traditional two-dimensional images have problems such as information loss.Therefore,people use tools such as 3D lidar sensors to establish a 3D target detection system,including indoor and outdoor detection.This paper takes the current popular autonomous driving field as an entry point,based on 3D object detection,among which point cloud-based methods are dominant,which can be further divided into multi-view-based,voxel-based,point-based and based on representation learning.Point voxel-based methods.This paper mainly focuses on voxel-based methods.There is an easily overlooked but very important problem in 3D object detection in autonomous driving,that is,the sparseness of point clouds,which will lead to the loss of a large amount of object information and spatial information.Different from ordinary 2D images that contain semantic information,3D point clouds are just the opposite in this respect but have a lot of geometric information,because there will be more 3D point clouds around vehicles or pedestrians.But due to the properties of point clouds,these clusters are unevenly sparsely distributed.Therefore,the purpose of this paper is to convert sparse point clouds into dense point clouds and complete the semantic information of point clouds,so as to improve the detection performance of small and medium-sized targets and medium-to-long-distance targets in 3D target detection in a data-enhanced manner.In terms of method,it mainly follows the idea of3 D super-resolution method of deep learning to reconstruct the dense point cloud.Compared with traditional point cloud densification methods,the Li DAR super-resolution algorithm based on deep learning proposed in this paper has achieved remarkable results on 16-channel self-collected point clouds,Point Cloud++ 16-channel and KITTI datasets,which are better than traditional methods.In the third chapter,a large number of visual comparisons are carried out.On this basis,this paper further discusses whether high-quality densified representations are helpful for 3D object detection methods.In this regard,a 3D object detection model based on voxel representation is selected as the backbone network,and its generalization is tested from various aspects.The generalizability and universality,including supervised and semi-supervised tasks,have achieved significant improvement on the KITTI dataset and ONCE public dataset,significantly outperforming existing baseline methods.Through a large number of visual comparisons,it is concluded that the proposed The deep learning-based Li DAR super-resolution algorithm can be used as a plug-and-play module to improve the detection performance of small and medium-sized targets and medium-to-long-distance targets in 3D target detection by means of data enhancement.