6D Pose Estimation Of Ship Object Based On Point Cloud Data

6D pose estimation is an important research branch of computer vision technology.The pose of an object is composed of three-dimensional position and three-dimensional direction angle.Especially in the field of ship awareness,intelligence has become the main direction of ship development,and the awareness of ship navigation situation is the key link of intelligent ship technology.For maritime visual monitoring,with the active maritime activities,it is difficult for manual monitoring to objectively judge the ship’s grade,intention and state.Therefore,the intelligent visual monitoring system which can provide situation awareness,threat assessment and even auxiliary decision-making is of great significance in maritime management and monitoring tasks.6D pose estimation of ship object is the key link of intelligent sensing system.Therefore,this paper develops 6D pose estimation of ship target.For the point cloud scene of ship object,an end-to-end 6D pose estimation algorithm is proposed to estimate the 6D pose of ship target.Firstly,a dataset of marine ship target point cloud is constructed.This paper summarizes the current ship datasets which can be used in the field of deep learning.These datasets are based on RGB images and lack of training tags for 6D pose.However,the existing point cloud datasets do not contain ship point cloud data.In order to solve the problem that there is no open point cloud dataset for ships at present,based on the ROS(Robot Operating System)and Gazebo simulation platform,the point cloud dataset which can be used in the marine scene is constructed by ourselves,and the 6D pose annotation is carried out,which lays the foundation for the subsequent 6D pose estimation of ship targets.Then,the ship object feature extraction technology based on point cloud data is studied.Based on the research of Voxel Net voxelization algorithm and Point Net++ algorithm principle,Point Net++ is improved.In the network structure,four Set abstraction layers are used to extract the features of the original point cloud,and the original point cloud is encoded as a vector representation with high-dimensional features.Voxel Net,Point Net++ and improved Point Net++ are tested in self built ship point cloud dataset and public dataset Model Net40.The experimental results show that the improved Point Net++ feature extraction algorithm is effective.Secondly,an end-to-end 6D pose estimation algorithm is designed.The algorithm uses the improved Point Net++ point cloud feature extraction network as the backbone network,designs a new pose regression network branch on the basis of 3D object detection Point RCNN,fuses the 3D frame of Point RCNN,and outputs 6D pose estimation results including 3D angle.In addition,the loss function of the algorithm is designed,and a new 6D pose estimation network is trained to realize the 6-DOF pose estimation of the ship object.Finally,the comprehensive experiment and analysis of 6D pose estimation of point cloud are completed.Based on the self-built ship point cloud dataset,the end-to-end 6D pose estimation network proposed in this paper is trained,and on this basis,the accurate estimation of ship pose in different marine scenes is realized.The algorithm is compared with pointrcnn and clustering algorithm,and the superiority of the algorithm is verified;And based on the public YCB video dataset,the proposed algorithm is compared with five other 6D pose estimation algorithms based on deep learning.The experimental results show that the proposed algorithm is superior to other algorithms in terms of estimation accuracy and algorithm complexity.