Research On Multi-view Stereo Reconstruction Method Based On Deep Learning

3D reconstruction technology has been widely used in fields such as autonomous driving,robotics,virtual reality,and has attracted increasing attention in recent years.Despite many advanced 3D sensors,such as LiDAR,are widely used in some realtime 3D reconstruction tasks such as Simultaneous Localization and Mapping(SLAM),scenes reconstructed using sensor-based methods are usually sparse and the high cost of these precision sensors results in a high reconstruction cost for sensor-based method.For tasks that do not require real-time performance,image-based methods are often adopted,which only use sequential images.These methods obtain depth information by matching 2D images,and the entire process is completed offline.Compared to sensor-based methods,image-based methods are more economical and practical,and the resulting 3D models are more detailed and complete.The depth-map-based method is the simplest and most flexible way of image-based methods,which mainly includes two key steps:camera pose estimation and depth map estimation.Camera pose estimation mainly relies on image feature matching technology,while depth map estimation mainly utilizes multi-view stereo technology.Therefore,this thesis starts from these two key technologies.Two algorithms are proposed respectively to address the problems existing in these two technologies for improving the final reconstruction performance.For multi-view stereo technology,this thesis proposes an interval-aware epipolarguided multi-view stereo algorithm,in which the epipolar-guided convolution can align the receptive fields of matching points between different views,and the features obtained in this way can construct a more accurate and robust cost volume;the interval-aware depth prediction module combines the advantages of previous methods,which can not only directly supervise the cost volume but also locate the correct depth interval to avoid interference from irrelevant depth hypotheses.This thesis conducts experiments on multiple image feature matching and multiview stereo datasets,and the experimental results show that the two proposed algorithms surpass other advanced methods,and the ablation experiments also prove the effectiveness of each module in our algorithm.Finally,we combine the two proposed algorithms to obtain good 3D reconstruction results.