弱特征表面形貌测量的低成本渐进式匹配方法

Low-cost incremental registration method for measuring the surface topography of weak features

弱特征连续表面的匹配问题是计算机视觉和图像处理中的一个挑战性问题,基于标准夹具的特征辅助可以准确地实现测量数据与模型的匹配,但该方法成本较高。针对该问题,提出了一种低成本的渐进式的匹配算法,首先基于带有边界罚函数的“点到面”ICP算法实现粗配准,接着通过对测量数据进行两步微调即可对标工业常用评价结果。以汽车玻璃的面形误差评价为例,仿真和实验结果表明,基于所提方法匹配后的面形误差接近传感器本身的误差级别。对于40 cm×40 cm的汽车玻璃,基于所提方法匹配线结构光测量的玻璃数据与三坐标数据,两者偏差在-0.06/0.08 mm,基本满足工业需求。

Objective The registration problem of continuous surfaces with weak features poses a challenge in computer vision and image processing. Automotive glass is a typical example of such a continuous surface with weak features. Due to the unclear characteristics of the Reference Point System(RPS) on automotive glass, it is difficult to accurately register the three-dimensional data obtained from either three-coordinate measurement or optical methods to the RPS coordinate system. To address this issue, the research proposes a low-cost and progressive registration algorithm that does not rely on high-precision fixtures and can still achieve precise registration and dimensional evaluation.Methods This method first builds upon the "point to surface" ICP registration, and further proposes a rough matching with boundary penalty function correction to achieve initial alignment between the measurement data and the model, providing good initial values for subsequent registration(Fig.1). Secondly, in order to align with the results of CMM's measurement and meet industrial needs, the distance between non-RPS sampling model points and corresponding measurement data points is directly optimized through CMM's evaluation method, and the measurement point cloud is fine tuned. In order to adjust the overall minimum registration benchmark from the previous step to the RPS benchmark(Fig.2), the distance between the RPS point and the corresponding point in the measurement data was directly optimized using CMM's evaluation method, thereby achieving accurate positioning and surface shape evaluation of the RPS point in the measurement data of weak feature continuous surfaces.Results and Discussions This article validates the effectiveness of the proposed progressive RPS point positioning algorithm through simulation and actual experiments. Combined with(Fig.11), it can be clearly seen that each step of adjusting the measurement data makes the deviation between the measurement data and the model closer to the deviation between the three coordinates and the model. Finally, the surface error between the measurement data and the three coordinates is controlled at the level of-0.06/0.08 mm, as shown in Fig.11,slightly greater than the measurement error of the sensor. The results are basically consistent with the simulation results, as shown in Fig.6, which meets the actual requirements. Because each step of registration is based on sampling points, which are uniformly sampled based on the model, although there are a total of three steps, fast registration of 200 000 measurement point clouds can be achieved in about 10 s using the RPS point coordinate system. And from Fig.11, it can be seen that without any step, the matching error between the measurement data and the three coordinate measurement results is much greater than the deviation between the complete progressive registration measurement data and the three coordinate measurement results. This indicates that each step of the proposed method cannot be ignored.Conclusions This article proposes a low-cost progressive registration algorithm to overcome the technical difficulties of optical measurement methods in measuring weak feature surface RPS localization. Experiments have shown that for 40 cm × 40 cm automotive glass, traditional methods based on point-to-point ICP matching or direct matching with models have significant positioning deviations for RPS points, resulting in a significant discrepancy from the evaluation results of the three-coordinate system. However, the deviation between the proposed method and the three-dimensional evaluation is-0.06/0.08 mm, which basically meets the industrial demand, indicating the progressiveness of the proposed method. In addition, the proposed method actually has high computational efficiency. In the future, on the basis of improving the accuracy of automotive glass surface data, it is expected to combine the method proposed in this article to achieve online full inspection of automotive glass.