The Study Of Feature Extraction And Matching Technology For 3D Reconstruction

In recent years, three-dimensional reconstruction technology is widely used in some applications such as archeology, architecture, geology, virtual reality, robot navigation, object recognition and military fields. Meanwhile, with the rapid development of computer technology and computer network, the 3D reconstruction can be seen in all aspects of life and also has become the new hotpot, the 3D reconstruction is the process converting two-dimensional structure information such as the images or videos to real three-dimensional model by the computers.Feature extraction and matching is the first and vital step of 3D reconstruction, and it is the base of Euclidean reconstruction and camera self-calibration etc. The precision of image matching directly affects the results of following steps in the whole process. This paper focuses on two problems of feature extraction and matching. One is the distinctiveness of descriptors, and the other is the strategy of mismatching elimination for sequence images.To improve the distinctiveness of descriptor, in this paper, a new descriptor based on circle-blocks neighborhood division and normalized color illumination invariant (CCD) is presented. First, it adopts circle-blocks structure to divide neighborhood of feature points into 24 non-overlapping sub-regions in the polar coordinate space, and then gives weight calculation method for evaluating the effect of every sub-region according to the distance between every sub-region and the feature point. Next, the illumination invariant is formed by the normalized RGB triple channels color intensity differences. Finally, a 72-dimensional feature vector will be established by combining the weights with illumination invariant.To solve the problem of mismatching caused by the large number of similar structures in sequence images, this paper presents a mismatching points elimination algorithm based on the geometric constraint conditions for sequence images. The mismatching points are eliminated according to the relationship of consistent motion orientations and equivalent motion distances between them in the same scale. And the mismatching points are eliminated according to the intersection relationship of lines between them in different scales.Experimental results on real images demonstrated that the new algorithms proposed in this paper are valid and robust under various geometric and photometric transformations application such as scale changes, viewpoint changes, translation transforms etc.