Research On 3D Reconstruction Algorithm Based On Multi-view Stereo

3D reconstruction is a frontier and popular direction in computer vision and computer graphics with a wide range of applications,including medical research,heritage restoration,intelligent transportation,virtual reality,etc.Compared with 2D images,3D models have more three-dimensional and vivid visual effects.Traditional active vision methods use hardware devices such as scanners to obtain 3D models with high accuracy but at high cost.Since high-resolution images are easier to obtain,researchers have focused more on reconstructing scenes using captured images.Compared with traditional active vision 3D reconstruction methods,optical image-based 3D reconstruction methods improve reconstruction accuracy while reducing reconstruction costs.The optical image based 3D reconstruction is divided into sparse reconstruction and dense reconstruction,and this paper focuses on the dense reconstruction based on Multi-View Stereo(MVS),which is a dense reconstruction of the scene using different viewpoint scene images,whose input is a calibration image set and sparse feature points,and the output is a 3D point cloud model.Although the existing 3D reconstruction methods reconstruct the model with good visual effect,the accuracy and completeness of the model still need to be improved.Compared with the existing methods,the main innovations of this paper include the following two points.First,a pixel-level stereo matching depth map estimation method based on local a priori is proposed.To solve the problems of low accuracy of map selection,low computational efficiency,and slow convergence of depth values,three main studies are conducted as follows: first,to improve computational efficiency,the smallest image set that can cover the whole scene is selected as the reference image set,and a suitable neighboring image is selected for each reference image.Second,to accelerate the depth map convergence rate,the depth values are initialized using local feature points of pixels and a list of candidate image indexes is initialized for each pixel of the reference image for dense matching.Finally,the matching cost is calculated using an improved normalized cross correlation function to solve the depth discontinuity problem in the local region,and the matching cost with the highest score is used as an indicator of the propagation in the horizontal and vertical planes to iteratively update the depth values.Experiments show that the method is relatively better than other methods in terms of accuracy and speed.Second,a multi-scale depth map estimation method based on geometric constraints is proposed.To solve the problem of low reconstruction completeness of weak texture regions in the scene,low size images are used to guide high size images to fill the depth information of weak texture regions.Based on the advantage that the smaller the image size is the richer the texture information contained in the matching window when the matching window size is constant,the depth map of the smallest size reference image is first computed using a local a priori pixel-level stereo matching depth map estimation method,then the depth map is upsampled and a geometric constraint of front-to-back reprojection is introduced to reduce the depth error when the image is upsampled,and finally the depth map optimization and depth map fusion are used to The 3D point cloud model is generated.Experiments show that the method outperforms other methods in terms of completeness and the reconstruction results are better overall.