| Budget Amount *help |
¥3,700,000 (Direct Cost: ¥3,700,000)
Fiscal Year 2004: ¥1,800,000 (Direct Cost: ¥1,800,000)
Fiscal Year 2003: ¥1,900,000 (Direct Cost: ¥1,900,000)
|
|---|
| Research Abstract |
As non-rigid objects, (1)water, (2)botanical tree, and (3)human images are dealt with by this research. Specific studies of recognizing each non-rigid object's behaviors by computer vision technologies and reproducing in 3-dimensional virtual space are described as fellows.
(1)A method for analyzing the video images acquired by a camera that observes the surface of water so that the shape of water surface is reconstructed by Bump mapping is studied. To perform Bump mapping, it is necessary to estimate the surface normals to many points in the water surface, and Shape from Shading(SfS) is considered to be useful for this purpose. To suppress reconstruction errors caused by SfS's orthographic projection camera model, a method that optimizes the obtained orientations of the surface normals by a genetic algorithm is proposed. Some promising experimental results were obtained.
(2)In order to reproduce real botanical trees' behaviors in a 3-dimensional tree model, it is necessary to recognize
… More
the real trees' behaviors, but it is very difficult to track individual leaves and branches due to occlusions. Therefore, a method that stores video sequences that contain the trees' behaviors caused by different winds and estimates the direction and strength of the wind by a learning procedure that utilizes the stored data is proposed. Experimental results show the effectiveness of the proposed method.
(3)A method for estimating a human body's postures by a computer vision technology is studied. A human who wears multiple-colored suit, in which each body part is colored differently, is observed by cameras, and each body part is extracted by color information processing : i.e., among the multiple images, two regions having the largest and 2^<nd> largest numbers of pixels are selected, and the shapes of the two regions are analyzed so that the 3-dimensional coordinates of joints are obtained. The effectiveness of this method was confirmed experimentally. Since the above-mentioned computer vision based method gives worse estimation results for the accuracy (resolution) and the number of degrees of freedom with respect to the posture than contact-type motion capture systems, a method that solves this problem by exploiting a learning function by neural networks is proposed, and its effectiveness was confirmed experimentally. Concerning facial expression reproduction in a 3-dimensional face model, a method that maps facial textures extracted from a real face video sequence to the deformed face model was applied to representing precise structures such as wrinkles realistically. Less
|
