Design And Implementation Of 3D Scene Semantic Mapping Algorithm Under Space On-Orbit Service Task

Space on-orbit service is an important guarantee to improve the operation reliability and prolong the service life of spacecraft.The failure of satellite solar panel deployment mechanism is the most common space satellite failure,which is mainly due to the incomplete action of explosion bolts.At this time,relevant cutting and other operations shall be carried out on the unexploded or incompletely exploded bolts through the mechanical arm carried on the space service satellite.In order to realize the environmental perception and obstacle avoidance in the above tasks,a 3D scene semantic mapping system is designed in this thesis.The system senses the obstacle information and target object coordinates in the environment through the visual sensor at the end of the robot arm,and constructs a 3D semantic map to assist the robot arm serving the satellite to complete the task of approaching the target bolt and cutting.The implementation of the system is mainly divided into two problems: semantic segmentation and mapping.For these two parts,an optimization algorithm of explosive bolt semantic segmentation based on attention mechanism and an object level 3D semantic map construction algorithm are designed.Aiming at the situation that the target object is blocked or the target object is small in a long distance in the practical application process,the explosive bolt semantic segmentation algorithm based on the attention mechanism adds the fusion attention module between each convolution layer of ResNet.ResNet is the backbone network of semantic segmentation model MaskRCNN.The fusion attention module is composed of spatial attention module and channel attention module in series to optimize the attention mechanism of the feature image output from the convolution layer.The algorithm improves the segmentation accuracy and robustness of the model.Finally,the semantic segmentation model is trained and its results are tested.Object-level 3D semantic map construction algorithm aims to achieve higher-level semantic map.Objects in the map are stored as independent individuals and retain their own attributes and geometric models.The 3d semantic map construction algorithm mainly integrates semantic information into the traditional visual SLAM algorithm,and constructs a dense map of the environment based on the camera pose estimation provided by the visual odometer.In this paper,the 3D semantic map algorithm is gradually implemented in four parts: temporary object mapping,point cloud registration,object association and 3D mapping.Aiming at the problem of semantic missegmentation of bolt edge pixels,a bilateral filtering algorithm based on feature domain was designed to remove the edge background noise generated after temporary object segmentation.The generated temporary objects are matched with historical objects in the environment through point cloud registration algorithm.In this paper,an improved ICP algorithm based on photometric error is designed by applying the mapping relationship between point cloud and color images in SLAM system,which improves the matching efficiency and accuracy.In the data association part,corresponding data association and model update methods are designed respectively on the basis of SLAM system to realize the real-time maintenance of object database.Finally,an octree graph building algorithm based on fast rasterization is designed to improve the efficiency of graph building.In order to verify the validity of 3D semantic mapping system,semantic segmentation and mapping are tested respectively.The semantic segmentation of explosive bolts is verified by bolt morphology and illumination conditions.The results show that the model has strong robustness for bolts of different shapes,but it is necessary to use light filling device to stabilize the light source in complex and changeable lighting environment.The validation of the mapping algorithm includes the validation of the object data set and the validation of the global semantic map.The results show that the semantic mapping algorithm can truly annotate the object information in the environment,and the mapping speed of the octree mapping algorithm based on fast rasterization is increased by 200%.