Study On Image-based Modeling For Plants

Image-based modeling (IBM) is a research topic in Virtual Reality field. Its main objective is to capture the inner and outer parameters of the camera so as to reconstruct the 3D model of the object by processing the image information taken by the camera. Compared to traditional 3D modeling methods, such as 3D modeling software and laser scanners, IBM has the advantages of low cost and high level of automation, and considering the reconstructed model, it is complied with the topology of the original object. How to reconstruct the 3D information from noised or incomplete 2D information is a big challenge for computer vision, and in this paper, we focus on this challenge as our research target.Considering our object - plant, we adopt the method which is suitable for the uniform-textured object like plant: IBM from the active cues, weak-structured light. The "weak-structured light" refers to the cast shadow on the object when the operator holds a stick and moves above the plant.We design a prototype system based on the active cues "weak-structured light", and divide our system into five parts, which are: 1. the calibration of the inner and outer parameters of the camera and the calibration of the light source position; 2. the calculation of the shadow plane; 3. the calculation of 3D information from the 2D coordinates; 4. the registration of multi-views data; 5. the triangulation of the unorganized points. We also implement this system by integrating the above five parts, and only a few interactions is needed before the reconstruction of 3D model through the calculation of weak-structured light from multi-views. Our main contributions are:In the shadow plane calculation part: how to simulate the projective line of the intersection of the shadow plane and the reference plane is the crucial step in recovering the shadow plane. In order to implement the high accuracy and automation simulation, we raise a new algorithm to simulate the line based on the least point-line square distance. The method we raise not only solves the error the least-square method brings, but also implements the accurate simulation in the complete automatic calculation.In 3D reconstruction part: We raise a distance square sum based algorithm to automatically simulate the 2D line, which solves the Least Sqaure Method's error when the line turns nearly vertical to X axis. Considering noises in the frames, we propose many steps to largely reduce the noise influence.In the modeling part: based on the topology of plants, we raise two-step registration method. The multi-view registration will somehow lead to the multi-layer of the same leaf. We use the unorganized points of each leaf to calculate the control points of B-spline surface and generate the B-spline surface based on these control points. This method not only solves the multi-layer problem, but also fines the point coordinates based on multi-view information.