基于多目立体视觉和神经网络标定的表面形貌测量方法研究

针对三维表面形貌的非接触式光学测量是计算机多目立体视觉技术的一项重要应用，但目前还存在相机个数受限、特征点匹配算法复杂与纵向测量精度不够等难题。开发了一种基于多目立体视觉和神经网络标定的表面形貌测量方法，其中包括:使用神经网络完成多目标定与三维重构，在表面投射激光点阵作为图像识别与匹配的特征点，应用蚁群粒子跟踪测速技术进行多相机间相同特征点的匹配。经实验测试，相较于传统基于小孔成像模型进行标定与基于核线约束或互相关算法进行匹配的立体视觉测量系统，所提出的方法可适配具有大光学畸变的场景，能有效提高测量的空间分辨率，深度方向的测量误差在1.0%～2.0%的水平。

Research on surface measurement method based on multi-view stereo vision and neutral network calibration

Non-contact optical measurement of three-dimensional surface topography is an important application of the computer multi-view stereo vision technology. However, there are still some technical problems needed to be solved such as limited number of cameras, complex image matching algorithms and insufficient longitudinal measurement accuracy. The article develops an optical method for surface topography measurement based on multi-view stereo vision and neural network calibration. This algorithm includes: applying neural networks for the calibration of the multi-view cameras and the three-dimensional reconstruction of feature points, projecting a laser speckle pattern on the surface as the feature points for image recognition and matching, using the particle tracking velocity technology to match the same single feature point in the multi-view images. As proved by experiments, compared with the traditional stereo vision measurement system, which is calibrated based on the pinhole camera model and matched by the epipolar constraint or cross-correlation algorithm, the method proposed in this paper can be adapted to scenes with large optical distortion, besides, the spatial resolution of the surface topography measurement is effectively improved and the measurement error in the depth direction is at the level of 1.0% – 2.0%.