A Study On Three Dimensional Reconstruction Based On Projection Of Virtual Height Line

The key of exploring environment is sensing and recognizing three dimensional environment and using the reconstructed results in robot navigation, industrial manufacturing and testing, virtual reality. Recently, with the rapid development and wide use of three dimensional reconstruction, many applications have put forward requirements of achieving the scene information accurately. As one of important techniques of three dimensional reconstruction, stereo vision has been an important research, because of its characteristics of accurate and fast measurement. In general, the traditional three dimensional reconstruction method reconstructs the scene by computing the three dimensional coordinate of each pixel in dense disparity map obtained by stereo matching. It usually requires high computing ability and costs a lot of time, and large redundant data exist in the reconstructed data. For the problems existed in the traditional method, this paper proposes a three dimensional reconstruction method based on projection of virtual height line(PVHL).Firstly, the algorithms of three dimensional reconstruction are intruduced including camera calibration, image rectification and stereo matching. The camera's interior and exterior parameters are acquired by camera calibration. Using image rectification, the corresponding epipolar lines are made parallel to the horizontal scan lines of image pairs. Combining the proposed three dimensional reconstruction model, the appropriate stereo matching algorithm is selected.Then the three dimensional reconstruction method based on PVHL is introduced in detail. Firstly, the scene was partitioned into grids uniformly, and then virtual height line segments(VHLS) were introduced in the grids and projected into the stereo images to generate projection line segments(PLS) in stereo vision system. Finding the height of the scene point was converted to prove whether the similarity metric of the projected point has the maximum similarity metric on the VHLS and in the searched horizontal disparity range or not. The candidate correspondences were obtained under the condition that the requirement of the maximum similarity metric on the VHLS was relaxed in order to reduce the impact of quantization error of PLS and window effect of local matching. Then the candidates were refined by edge features. Finally, tabulating method and box-filtering were adopted to decrease redundant compute, and hyper-threading was used to shorten the runtime. The experiments to reconstruct 3D scene are performed and the results show the proposed method is effective, more accurate than the traditional method, and meets the requirement of real-time.Spatial resolution and image resolution, which have a greater impact on the results of three dimensional reconstruction, are studied. In the case of different spatial resolution, the qualitative and quantitative analyses have been made on the reconstruction results of artificial scenes and natural scenes. And the method based on the hierarchical level interest grids reconstruction is proposed, by which more scene detail and more accurate results could be obtained. Experiments have been made on the scene images with different image resolution. The experimental results show high image resolution could reduce the influence of image quantization to the location of PVHL and the reconstruction error.Finally, the error analysis of the proposed three dimensional reconstruction method has been done, and the error compensation model has been established. The experimental results of error compensation show that the compensation method can reduce the reconstructed error and make the reconstruction results more precise.