Research On 3D Reconstruction Technology Based On Mobile Imaging Sensor

3D reconstruction based on unordered images acquired by mobile imaging sensor is a popular research direction in computer vision.This technique has been used in many applications,such as quality control system,digital heritage preservation,digital mapping and navigation,mixed reality,and many more.The objective of 3D reconstruction based on unordered images is recovering scene structure and camera parameters from image data,which can also be called Structure from Motion.This thesis studied image-based 3D reconstruction system and in order to improve the computation efficiency of feature matching and geometric verification,methods were presented in this stage,which finally improved result of 3D reconstruction.The main work of this thesis is organized as follows:1.An approach was proposed to efficiently decide on whether an image pair has scene overlap or not based on image features.This approach is applied to avoid useless computation between feature extraction and feature matching.Firstly,mapping features of two images to a descriptor separately and then establish the relationship between two images according to the descriptor.Finally,a learning-based model was trained to predict scene overlap or not between all the image pairs.Experiments indicate that this approach can efficiently reduce the computational complexity in feature matching.2.An approach was proposed to improve the efficiency of geometric verification based on the result of feature matching.There are some outliers in feature matching results due to measure noises or observation errors.The efficiency becomes lower while the outlier ratio is higher during geometric verification.This thesis constructs a classification feature vector based on the putative correspondences in feature matching and then perform outlier detection based on support vector machine so that part of outliers can be removed.Experiments indicate that this approach can effectively reduce the outlier ratio and improve efficiency of geometric verification.3.An unordered image-based 3D reconstruction pipeline was established.Based on thorough research and analysis on all the stages of 3D reconstruction,a complete reconstruction system was established,which including the proposed approaches and feature extraction and matching,sparse reconstruction,dense reconstruction,etc.