Measurement of Automobile Frame Circular Holes Based on Point Cloud Completion

The point cloud data of the automotive frame collected by a laser scanner can achieve very high precision when measuring circular holes on the frame. However, the point cloud data is incomplete because of occlusion and limited resolution, which affects the accuracy of the measurement results. Therefore, partial point cloud completion is crucial for improving the measurement accuracy of large-sized parts. This paper investigates a method for the completion of incomplete point clouds of automotive frame parts based on deep learning neural networks, and extracts the circular hole features of the frame from the completed point clouds while estimating the feature parameters of the circular holes.Firstly, the point cloud is encoded with positional information and transformed into a sequence of point proxies. A Transformer encoder-decoder structure is employed to complete the point cloud. Then, the plane fitting method combining Random Sample Consensus (RANSAC) with density-based clustering is improved to detect circular hole planar features from the completed point cloud and. The least square method is ultimately applied to fit the circle and estimate the radius and center of the circular hole. The experimental results show that the measurement accuracy error of the circular hole radius obtained by this method is less than 0.5 mm, and the repeatability error of the circular hole center positioning is less than 0.025 mm.