Three Dimensional Shape Modeling by Multiple Sensor Information

发展了一种新型的人工神经网络，用于三维形状的自动建模。这些神经网络具有从其多个视点图像学习建模三维形状的功能的能力。为了对形状进行建模，他们使用多面体、由网络的特殊架构表示的模型以及网络之间的连接权重。在学习过程中，通过使用多个视图图像来更新这些连接权重，使得多面体模型的视图与实际形状的对应视图一致。通过实验对所提方法的性能进行了评估。该方法结合红外照明法测得的多视点图像和其他距离信息，取得了较好的效果，对噪声和误差具有较强的鲁棒性。该方法还联合应用了基于立体匹配的测距方法，并通过误差修正修正了相互矛盾的结果。

Novel Artificial Neural Networks were developed for automatic modeling of three dimensional shapes. These neural networks have capabilities of learning a function of modeling three dimensional shapes from their multiple view images. To model the shapes, they use a polyhedron, models represented by the special architecture of the network and connection weights among the network. These connection weights are updated during the learning process by using multiple view images so that the views of the polyhedron models coincide with the corresponding views of the actual shapes. The performance of the proposed methods is evaluated by some experiments. By combining multiple view images in addition to other distance information measured, by infra red illumination method, the method gave improve results with robustness against noises and errors. Stereo matching based distance measurement method is also jointly applied to the method and contradicting results were fixed by error correction procedure.