基于结构照明和BP神经网络的三维物体识别

Rotation-invariant 3-D object recognition based on structured light illumination and BP neural networks

提出一种具有旋转不变性的三维物体识别的新方法,该方法通过结构光照明的方法,使物体的高度分布以变形条纹的形式编码于二维强度图中,由于条纹图包含有物体的高度分布信息,因此对条纹的相关识别具有本征三维识别的特点。旋转不变性是通过BP神经网络实现的。计算机模拟结果表明,用二维强度像的基频分量做训练样本设计BP神经网络,选择训练样本和隐藏层神经元的数目,基于结构光编码的BP神经网络对三维物体具有良好的旋转不变识别效果。

An new method for rotation-invariant three-dimensional （3-D） object recognition was proposed. The method was based on the use of 2-D information encoded in the form of deformed fringe pattern which was obtained when a grating was projected onto an object＇s surface. The deformed fringe patterns contained the height information about the objects, so the method had the characteristic of intrinsic 3-D recognition. The rotation invariance was achieved by BP neural networks. Through designing BP neural networks with the first-order frequency as training samples and selecting the number of training samples and hiding neurons, the results of computer simulation show that BP neural networks combined with structured light illumination has a better performance for rotation-invariant 3D object recognition.