Parallelization Research And Implementation Of Obtain And Use On Key Elements In 3D Reconstruction

With the development of artificial intelligence technology,three-dimensional reconstruction technology as the support of the autonomous driving and virtual reality applications are widely concerned.However,the high computational complexity and the huge data processing scale of the application program have restricted the real-time capability of the three-dimensional reconstruction.For this reason,the following specific research is carried out for several typical scenes,reconstruction methods which are commonly used in the following processes:(1)In the small indoor scene reconstruction,there are many excellent feature detection algorithms such as Fast,Surf,SIFT etc.In recent years,the development of SIFT is relatively mature,and SIFT is adapted to obtaining the key elements in the scene.This part has been optimized by cuda parallel technology,which is more time-consuming in the matching process based on feature extraction and calculation.Simultaneously,in the phase of feature matching,a KD-Tree search method based on GPU is implemented,and the feature matching process is accelerated by KD-Tree and the cache keys which have the red-black tree feature.(2)In the course of an ICP registration,the region of interested optimization KinectFusion algorithm is introduced,the method is verified by experiments and overall computing performance has been effectively improved,under the premise of guaranteeing the reconstruction precision by analyzing the widely used Kinect sensors and the rapid modeling of KinectFusion provided by Microsoft in sports scenes.(3)In the process of large scene reconstruction.With the rapid increase of computation,a parallel bundle adjustment algorithm based on GPU is achieved for the computational bottleneck of structure from motion for high speed imaging.Then,the sparse 3D point which used the structure from motion is transformed into a dense 3D point cloud by CMVS and PMVS technology.Next,a dense smooth model was constructed by parallel Poisson surface reconstruction method.Finally,the distributed three-dimensional reconstruction system is obtained based on the distributed framework,and do a series of system performance testing experiments.Experiments show that the distributed three-dimensional reconstruction system can effectively deal with the three-dimensional reconstruction of large scenes.And because of the introduction of CUDA parallel technology,the reconstruction efficiency improves obviously.However,the reconstruction based on the common camera(not depth camera)still can't meet the needs of real-time reconstruction.Secondly,the distributed reconstruction has high requirements on hardware configuration,and it doesn't apply to the field of mobile embedded terminal.Finally,we will analyze the incremental elements of the scene by means of deep learning,and carry out incremental reconstruction,on the basis of the existing model,thus reducing the computational complexity of reconstruction.