Research On Automation Technique For Close-Range Object's 3D Reconstruction

The paper's arguement and researches are mainly on how to take object's photographs by non-metric cameras and realize 3D reconstruction of true objects provided with control points. Level of automation of reconstruction for close-range objects is still low. The first reason is that close-range photogrammetry can not make the same normalized conditions as arieal photogrammetry and its basic photographing parameters's stableness. The second reason is that object's varieties and scene's complexity make it difficult for object's recognition and model representation. A system for image-based reconstruction is realized and several airplanes and chairs are reconstructed. The goal of this paper is to improve the automatic level of reconstruction in two respects of object extraction and modelling.The main works includes as follows:1. Method of photographing for close-range objects and orientation parameters solve. Here we decompose the object (airplane) to multiple strips and takes sequential pictures for every strip. So it is suitable to be processed for one strip using single strip analytical aerotriangulation. Firstly we can get interior parameters of the non-metric camera by DLT method based on available plane control field. Then for the exterior parameters of camera, we use two approaches: 1) for scene which is suitable for continuous orientation computation, we use the"strip triangulation"method for get the initial parameters of cameras. 2) for scene which is not suitable for continuous orientation computation, we use 3D-DLT and resection method for single image using enough control points. For these two approaches, Bundle Adjustment is implemented for get optimal results.2. Low-level features are basic primitives for image matching and measurement, are basic of recognition of middle, high features. The paper conducts corner points, line segments, region feature as feature primitives. These extraction's methods are argued, and these extraction algorithms are realized. It's work is basic of specific objects recognition.3. Automatic extraction for control marks. Mark is used widely in computer vision and Photogrammetry, especially for close-range for control points or texture feature fit for measure. So automatic extraction is needed for improve automation. But most extraction method is for simple background with little noise. In complex scene, manual measurement is required. We propose the method of automatic recognition and location based on object's invariant index. One Object can be presented by combination of low-level feature, which can be detected by existed methods. So we only use the aggregation of low-level feature detected as object's initial position in image. Then for specific object, it has own feature structure. We can design the specific algorithm to locate its key point as special feature point. For rectangle mark, we use template matching and line extraction for locate its center point. By this way we can automatically measure most mark points in every image.4. Full 3D Tin for object based on Delaunay TIN(D-TIN) Patch. For the terrain reconstruction, DEM and ortho-image are used to represent the terrain model. But for close-range objects, most are needed full space TIN. It is not sufficient to use just depth presentation. So we propose the TIN patch, which are defined by boundary and constraint D-TIN. The whole object's space TIN can be presented by different TIN Patchs which are combined seamlessly by connected boundaries. For every Patch TIN's generation, we can use correlation match based on epipolar constraint to measure surface's space points automatically because of the surface's spatial continuation for every Patch TIN. With texture mapping, we can get the whole 3D photo-realistic model of object.5. Some objects (plane's forward and backward circle parts,tyre,chairs) with regular geometric shape which are not suitable for TIN presentation because of high cost and coarseness, are presented by parametric geometric models and reconstructed based on multiple-views measurement. Objects, which are decomposed to simple geometric model easily, are reconstructed by using CSG method. The certain complex objects can be presented by generalized-cylinder integration method. The paper designs three kinds of cross-section(circles,Rectangle,Polygon) for objects with different geometric shape. Parametric geometric model's reconstruction includes restoration of pose and restoration of shape parameters. Several measurement approaches are designed and realized for model's pose measurement. Parametric geomeric model's adjustment can be conducted for these optimal solve. Shape parameters can be modified and restored by user's interactivity.