Research On The Optimal Algorithm Of 3D Reconstruction Based On Artificial Neural Network

Three-dimensional reconstruction technology has been widely used in the field of computer vision.The acquisition of 3D information from a two-dimensional image allows people to observe a scene from a more comprehensive and macroscopic perspective,and also enables the computer to have spatial awareness,which has greatly promoted the development of technologies such as unmanned and robot navigation.However,due to the diversity and complexity of the reconstruction scene,a large number of similar textures and background noise make the 3D reconstruction technology difficult to have both high precision and real-time at the same time,and raise the questions of enormous mismatches of feature points,insufficient accuracy of camera pose recovery and low global consistency of the reconstruction results.In order to achieve higher efficiency and high precision 3D reconstruction,this research proposes 3D reconstruction optimization algorithm based on artificial neural network,which uses the semantic recognition and nonlinear fitting characteristics of the neural network to optimize the feature point processing,pose recovery and loop closure.Specific work and innovation are as follows:Firstly,the nonlinear fitting ability of BP(Back Propagation)neural network was used to analyze the relationship between image complexity and the number of feature points,so as to adaptively adjust the number of detection points of feature points in different scenarios.The semantic segmentation characteristics of the FCN(Fully Convolutional Network)neural network were used to determine the main target range in the image to limit the detection area of the feature points and avoid the influence of background noise.To optimize the feature point matching and to obtain higher precision of the matching feature points,the stability of the camera pose shift during the 3D reconstruction process was analyzed and the pixel detection range of the feature points to be matched was limited.Then,the initial pose was calculated via the optimized matching feature points and the sparse point cloud was restored.The matching feature points,initial poses and sparse point clouds were later used as optimization variables to construct a maximum likelihood estimation problem.The corresponding re-projection error function was designed,and the error function is minimized through the graph optimization method.At the same time,the threshold culling error matching was set to eliminate the mismatches as well as to reduce the influence of the cumulative error,and the optimized camera pose was obtained.Finally,the traditional Bag of Words structure in loop closure detection was abolished.According to the semantic recognition result of scene image by Faster-RCNN neural network,a 2D semantic feature vector map was constructed,and a similarity based on nonlinear cumulative error was designed accordingly.Thereby,the loop closure could be identified,and a more accurate loop closure detection method was realized.