| Three-dimensional(3D)reconstruction is an important means to obtain 3D information of objects,and it is an important research content in intelligent perception,artificial intelligence and other fields.In recent years,3D reconstruction techniques based on structured light(SL)have made remarkable progress.However,the camera and projector must be synchronized in the reconstruction process.The removal of synchronization constraint is not only a difficult problem in the field of 3D reconstruction,but also an obstacle in the development of its theory and application.In view of the limitations of traditional synchronized SL 3D reconstruction models and the increasing demand of unsynchronized application scenarios,this thesis proposes an unsynchronized structured light(USL)3D reconstruction algorithm based on deep learning.For reference to the wide application of deep learning algorithm in image processing,this thesis proposes a deep neural network algorithm to realize unsynchronized aliasing pattern separation.Based on the characteristics of the aliasing pattern,the neural network structure is improved to learn the complex mapping relationship between the aliasing pattern and the associated projective pattern,so as to realize the high-precision aliasing pattern separation,accurately estimate the projective pattern intensity value from the aliasing pattern,and finally realize the unsynchronized 3D reconstruction.The main work of this thesis is divided into the following three parts:(1)Construct an imaging model of USLThe fundamental reason why the traditional SL 3D reconstruction model cannot handle the 3D reconstruction of unsynchronized system is that the 3D reconstruction model does not contain the camera transient imaging process and cannot describe the projection pattern changes during the camera exposure time.In this thesis,based on camera imaging principle and unsynchronized pattern timing analysis,an USL imaging model including transient imaging response is proposed to accurately describe the process of unsynchronized pattern imaging.(2)Aliasing pattern separation based on deep learningBased on the pattern aliasing mechanism and characteristics of unsynchronized systems,a deep learning algorithm is used to separate the intensity values of aliasing pattern images caused by unsynchronized systems.By making full use of the advantages of generative adversarial network,the generator structure and discriminator structure of the network are studied and designed respectively,and a multi-level adversarial network for aliasing pattern restoration is built to eliminate the influence of unsynchronized aliasing.The accuracy of aliasing separation is analyzed from two aspects of two-dimensional pattern image and 3D reconstructed point cloud.(3)Unsynchronized system design and constructionAn experimental system with controllable time parameters is established to precisely control the initial shooting time of the camera and the initial projection time of the projector,and a time-controllable USL 3D scanning system is realized.The visualization software system is developed,and the functions of pattern projection and shooting,system calibration,pattern decoding and 3D reconstruction are realized. |
