Projection-based registration using color and texture information for virtual environment generation

Abstract

In this paper, we propose a registration method for 3D data which uses color and texture data acquired from a multi-view camera for virtual environment generation. In general, most registration methods depend on 3D data acquired by precise optical sensors. However, as for a multi-view camera, depth errors are relatively large and depths in homogeneous areas are not measurable. We propose a projection-based registration method to cope with these limitations. First, we perform initial registration by establishing relationship between multi-view cameras through inter-camera calibration. Then, by applying color and texture descriptors to projected images, fine registration is accomplished. Finally, by exploiting adaptive search ranges, color selection is attained. Even if the accuracy of 3D data is relatively low, the proposed method can effectively register 3D data. In addition, an effective color selection can be done by setting up adaptive search ranges based on depth. Through this method, we can generate a virtual environment that supports user interaction or navigation.