The present work describes an industrial system for the automatic sorting of parts, from the customer’s requirements to its concrete implementation. The developed machine vision system is the central component of a packaging line which receives, as input, products made of polyurethane foam ranging in size from 5 cm × 5 cm × 5 cm to 70 cm × 70 cm × 70 cm. The purpose of the automatic packaging line is to increase the robustness and productivity of a process that has so far been performed manually.
The given problem is solved by registering a point cloud on the surface of the part. The point cloud is captured using an appropriate sensor system and then compared with reference point clouds obtained by sampling the original CAD models of the products expected on the conveyor belt. If the deviation of the points when comparing the captured cloud with a reference cloud is sufficiently small, we conclude that the currently observed part belongs to that reference model. Once the product has been identified, it can be packed into its designated box.
The results show that, using affordable hardware and the developed software, registration and subsequent product classification can be performed in real time. This is demonstrated on a challenging dataset of 24 products, 20 of which come in mirrored pairs. The required point accuracy for distinguishing mirrored parts of the specified sizes is 8 mm, i.e., an accuracy achievable by many depth cameras. By replacing the sensor with a more accurate one, the approach can easily be extended to quality inspection, since the key requirement for this is registration, which is already part of the architecture. The average decision time is approximately half a second, which, at the required conveyor belt speed of 6 cm/s, does not limit the productivity of the line.
|
