| Summary: | Three-dimensional (3D) geometric reconstruction and characterization of external anatomical structures from images has been a major research topic in Computer Vision. However, it is still a complex problem to solve, when automation, speed and precision are required. 3D human models are normally built using 3D scanners. Although frequently expensive, they are easy to use and can provide 3D models of great accuracy. Recently, volumetric methods have been successfully used in 3D reconstruction of objects with complex shapes. Comparing with the more usual stereo-based methods, they are more efficient in building 3D models when the objects involved have smooth surfaces. As they work in the objects volumetric space, their main advantage is to be free of any matching process between the images used. Matching is always very complex when smooth objects are considered as they do not have strong image features.Our work is based on Generalized Voxel Coloring (GVC) method, [1]. GVC does not impose any restriction on the objects shape or in the camera(s) motion. The images to be used in the reconstruction process are acquired by a simple fixed off-the-shelf CCD camera, which is calibrated using Zhangs method, [2]. Background/object segmentation is accomplished on all input images, in order to obtain the objects silhouettes. Having as inputs the objects image sequence, the associated silhouette images and the camera calibration parameters, the volumetric method GVC is applied. The obtained 3D model is then polygonized and smoothed, using Marching Cubes algorithm.For experimental evaluation purpose was used a hand and a human torso models. The camera calibration parameters were obtained with reasonable precision, and the 3D models built had satisfactory quality. However, it was verified that the reconstruction process is strongly affected by errors introduced by the camera calibration or by the segmentation processes. Thus, the future work will be regarding the improvement of these tasks. In the near future, some solutions to speedup the reconstruction process will be also addressed, as the development of a coarse-to-fine approach (e.g. using octrees) or the development of parallel implementations.
|
|---|