Key Technologies Of 3D Reconstruction Based On Portable Four-view Stereo

With the rapid development of computer technology,the demand for threedimensional(3D)information is increasing and 3D reconstruction technologies have attracted more and more attention.3D reconstruction technologies have the ability to obtain information about 3D size,geometric shape,surface texture and position distribution of the target,and are widely applied in cultural heritage protection,virtual/augmented reality,medical diagnosis,autonomous robot navigation,industrial automation,geographic information industry and other fields.The development of 3D reconstruction technologies contributes to the improvement of the automation and intelligence level of manufacturing,and offering much convenience to our lives.As an important research field of 3D reconstruction technology research,3D reconstruction based on binocular vision has attracted many researchers' attention due to its advantages of bionic hardware structure and low cost.However,binocular vision methods have the problems of low matching efficiency and unfeasibility for the scenes with weak texture.Additionally,binocular stereo matching in the areas with repeatable and similar texture features leads to ambiguous matches and the appearance of error 3D points.So binocular stereo matching methods have difficulties in reconstructing 3D models automatically.At the same time,only one scene with limited range can be reconstructed at a time due to the limited range of observation.In order to reconstruct a large scene,these small 3D models of several small scene from the division of the large one should be firstly generated.The reconstruction result can be obtained by the stitching of the small models from the overlapping scene.Aiming at the existing problems of binocular methods,a technical scheme for automatic 3D reconstruction based on portable four-view vision is proposed to achieve automatic 3D reconstruction of a single scene,automatic model stitching for several overlapping scenes and 3D measurement within the space range around 360 degree directions.The main contents of the thesis are as follows:(1)Aiming at the problem of the 3D dense reconstruction of the scenes with weak texture,a platform for four-view images data acquisition composed of a portable fourview system,a projector and a computer was built to capture the four-view images with projected patterns.Checkerboard patterns of array generated from the computer are projected towards the target scene via the projector for adding texture features to obtain a dense 3D point cloud.The density of the point cloud can be determined by the number of horizontal and vertical subdivisions in the checkerboard pattern array.By projecting checkerboard patterns,the problem of binocular vision methods' unfeasibility for the scenes with weak texture is solved.(2)Aiming at the problems of low binocular matching efficiency and ambiguity matches generated from traditional binocular stereo matching in the areas with repeatable and similar texture features,a four-view stereo matching algorithm is proposed to generate the 3D point cloud automatically from the four-view images.The algorithm uses multiple epipolar constraints in different directions from the four cameras and reversed self-tests of the matched points to reduce the mismatches.The algorithm also reduces the search range of matched points by using the intersection of epipolar lines to improve the matching efficiency.The experimental results verified that compared with the binocular matching,the four-view matching effectively reduces the mismatches and has less time consumption.(3)Aiming at the problem of much time consumption of triangulation of a point cloud in the 3D reconstruction process,a point cloud simplification algorithm is proposed to improve the triangulation speed of a point cloud and reduce the storage space of the 3D model by reducing the amount of 3D points in the point cloud.This algorithm can judge the flatness of the point cloud in each region from the division of the whole point cloud's region by plane fitting.Only the point cloud in the flat region is simplified.The amount of the points in a whole point cloud is reduced without loss of the details of local shape variance in the point cloud.The experimental results show that the algorithm obtained a good performance in the 3D reconstruction.(4)Aiming at the problem of 3D reconstruction of a large scene achieved by the stitching of several sub-scenes' models,an automatic model stitching algorithm based on SURF-ORB feature and space-distance consistency is proposed to transform the models of several sub-scenes to the same coordinate systems.Firstly,enough coarse matched points are extracted from the texture images of the overlapping scenes using the coarse matching method based on SURF-ORB feature.The correct matched points are chosen from the coarse matched ones through constant-distance consistency in the different coordinate systems using the fine matching method based on space-distance consistency.The accuracy of the model stitching can be ensured by the correct matched points.The coordinate transformation parameters are calculated from the correct matched points using least square method and the stitching of the overlapping scenes' models can be achieved.The experimental results of the stitching of six scenes' models show the effectiveness of the proposed method.The contrastive experiment between the proposed fine matching algorithm and the improved RANSAC algorithm was performed.The results of the contrastive experiment show that the proposed algorithm improves the efficiency of fine matching and the precision of model stitching compared with the improved RANSAC algorithm.(5)Aiming at the problem of the proposed stitching algorithm's unfeasibility for the model stitching between the scenes with weak texture,a panoramic four-view 3D measurement system was designed and implemented to achieve the 3D measurement of the space range around 360 degree directions.In order to unify the measurement coordinate systems of the five four-view systems in the panoramic measurement system into the same coordinate system,a panoramic calibration method based on rotation is proposed.The panoramic calibration of the panoramic measurement system can be achieved by building the virtual calibration field based on rotation.The experimental results show that the measurement relative error of the panoramic 3D measurement system is less than 0.2%.The experiment results verified that the panoramic 3D measurement system can offer a good hardware solution for model stitching in terms of measurement accuracy.In a word,the reconstruction of the single scene,automatic model stitching of overlapping scenes and other problems in 3D modeling are studied in this thesis.The software algorithm and hardware method of automatic model stitching are proposed.This thesis generally studies the process of automatic 3D reconstruction based on fourview stereo vision.