| With the rapid development of science and technology, human beings are becoming more and more inclined to apply machines and robots to deal with the daily work and solve some difficulty problems. So the robots are required to be able to observe and understand the world like human vision, and even adapt to the variety of the environment. Therefore, the computer vision comes.In the field of computer vision, how to extract the information from the 2D images acquired by a plurality of photosensitive elements to compose the appearance of the 3D object, is a kind of multi-view 3D reconstruction technique, and it’s research results has been applied in multiple field including medical imaging, topography detection, robot navigation, industrial parts measurement etc. However, the multi-view 3D reconstruction technology is not mature. Most of the applications of 3D reconstruction on the market are dependent on the expensive scanning equipment, which also limits the popularization of 3D reconstruction technology.In this paper, the related research of 3D reconstruction was studied based on multi-view stereo vision.The implementation method of the 3D reconstruction of multi-view stereo vision, was then analyzed. According to the existing defects in the present point cloud processing algorithms, a new point cloud simplification method was proposed. The main work in this paper was divided into the following parts:(1)A 3D reconstruction model was built based on multi-view vision. The camera imaging theory was studied. The classic imaging model and method was also summarized. Then the 3D reconstruction ideas in the each module of the three-dimensional reconstruction, were compared, such as the contrast between the corner and spot extracting method in the feature extraction,the contrast between feature points and feature region matching method in the stereo matching, the comparison of monocular, binocular, multi-view vision in the 3D reconstruction.(2)A new kind of camera placement array was proposed, which was annular multi-view placement array. After image acquisition from virtual 3D scene and image preprocessing, the ASIFT algorithm was utilized to extract features, match features. Then this array is used to adjust the epipolar line, calculate the 3D depth, reconstruct the 3D point, spread seed point, remove the dense point, and finally get the 3D point cloud of the scene. The experiments showed that the annular multi-view vision array could realize the 3D point cloud reconstruction of simple scene and complex real scene and obtained good results.(3)A simplified method of 3D point cloud reconstruction based on Snake algorithm was proposed. When the reconstruction area was chosen too big, it would lead to the target object unclear and influence the effect, which also took long time to reconstruct. Therefore, this method could simplify the reconstruction range through the snake contour extraction and finally reconstruct the object in the objective window. Experimental results showed that the proposed method could reduce the reconstruction time and improved the effectiveness of the reconstruction.(4)A point cloud mosaicing algorithm based on projection plane was proposed. When the reconstruction area was chosen too small, the target would not be complete, which would result in the lack of the useful information. Therefore,this method could mosaic the two-dimensional projection image of point cloud based on ASIFT algorithm, and then feed back the transform relation to 3D space, finally realize the point cloud and restore the complete information of the object. The experiment showed that this improved method compensated for the limitations and instability of the point cloud reconstruction. |
PDF Full Text Request