Comparative Analysis of Point Cloud Similarity Based on 3D Surface Reconstruction Using Mechanical Depth Sensor

Currently, significant progress has been achieved in the field of 3D reconstruction, and this trend is anticipated to persist for at least the next decade. In this research, we discuss the utilization of an Azure Kinect depth camera sensor to capture an image of a stamping die and a bottle, recording their depth point values for generating a corresponding point cloud for surface reconstruction. The height of each punch within the die is set manually using a hexagonal key to match the surface of the bottle. The procedural steps involve defining and extracting the field of interest from the original point cloud, incorporating additional filtering, such as passthrough and voxel, to eliminate undesired noise. To enhance processing efficiency during the point cloud registration, clusters are established within the point cloud to distinguish one punch from another and retrieve only the highest point of each. These peak points are then placed within the same coordinate system of the bottle's point cloud for further alignment and obtaining their fitness score at the convergence point. The average discrepancy in dimensions, measured in millimeters, between the actual object and the resulting point cloud is estimated to be less than 10%, with an average time of 3 to 4 min required for the overall surface reconstruction and point cloud registration.