Study On Reconstruction And Perception Algorithms Of 3D Scene

Reconstruction and perception of 3D environment serve as an important prior for automated systems,by providing task-relevant information about workspace.Related works aimed to build a framework for reconstruction and analysis of visual information presented in scenes,including category,position,and pose of objects.Due to the intrinsic large-scale and nonuniform characteristics,this topic encounters key issues such as robustness,efficiency,and feasibility.Accordingly,following contents were inspected:(1).Addressed for abstract descriptions of 3D image,feature extraction of point cloud was inspected,to improve the efficiency and robustness of existing feature descriptors,an enhanced feature descriptor equipped with exponential weights and bi-linear votes was proposed;(2).For compact representation of real world,fusion of multiple 3D images acquired from different views was investigated,feature-based matching and ICP method were combined for better stability;(3).An efficient dimension-reduced method was presented as an alternative to canonical feature-based algorithm for scene perception,which provides much better real-time performance.Moreover,a distributed structured point cloud protocol was designed with high extensibility,low redundancy,and ultra-rate I/O speed,which outperforms traditional point cloud in terms of flexibility and efficiency.(4).In combination of fully automated welding robots,comprehensive application of 3D scene reconstruction and perception was analyzed,including precise acquisition and modeling of 3D image,as well as configuration of sensor system in practice,efficient scene perception method was used to detect and extract object information.To integrate functionalities,a software platform was developed for simulation of 3D image related algorithms and application debugging.The contributions presented in this thesis have been validated by simulations and tests deployed in production environment.Featured contributions including:(1).enhanced feature descriptors outperforms existing methods on run-time,while maintains the discriminative power;(2).dimensional-reduced method can percepts object properties at real-time speed;(3).distributed structured point cloud supports real-time conversion,with ultra-fast processing speed that is distinctly better than traditional point cloud.