A Sparse Point Cloud Segmentation Method Based On 2D Image Driven

In recent years,with the increasing use of 3D sensors,3D point cloud segmentation tasks have become an important research topic in many applications of a large number of automatic navigation systems.However,sparse point cloud segmentation has become a very challenging topic due to sparsity,uneven sampling density,irregular formatting,and lack of color texture.Driven by excellent convolutional neural networks and large-scale datasets,2D image understanding tasks have made great progress,such as objection detection.Therefore,this paper uses the advantages of convolutional neural network in 2D images objection detection to assist the multi-line laser radar to complete the sparse point cloud segmentation task.This paper mainly focuses on the following research contents:1.The imaging model of monocular camera and the principle of 16-line lidar ranging model are introduced.A joint calibration model of monocular camera and multi-line lidar is established.Based on this,a joint calibration method based on plane features is adopted.The results of monocular camera calibration,monocular camera and 16-line laser radar joint calibration were analyzed.The effect of monocular camera and multi-line laser radar was visualized,and the feasibility of joint calibration method is verified.2.The principle of objection detection is introduced,and the difference between standard convolution method and depth separable convolution method is analyzed.An objection detection network based on depth separable convolution is proposed.For the objection detection network,the composition and principle of feature extraction network and multi-scale prediction network are analyzed.The definition of loss function is expounded,and the training process of the whole objection detection network is introduced.The experimental results show that,compared with the excellent YOLOv3 and Tiny_YOLO objection detection network,the proposed objection detection network has advantages in accuracy and speed,which verifies the feasibility and advantages of the objection detection network in this paper,and visualizes the effect of pedestrian detection,showing good results.3.A sparse point cloud segmentation method based on 2D image driving is proposed.The process of generating 2D image target candidate region based on the objection detection system proposed in this paper is described.The principle and implementation process of extracting the target 3D vertebral point cloud in the original laser point cloud are introduced.The process of selecting the final target point cloud in 3D frustum point cloud by designing threshold conditions using target category feature information is analyzed.The experimental results show that the proposed point cloud segmentation method has better accuracy and faster speed than the excellent point cloud segmentation algorithms such as Euclidean clustering and region growth.