Additive manufactured parts have complex geometries featuring high slope angles and occlusions that can be difficult or even impossible to measure; in this scenario, photogrammetry presents itself as an attractive, low-cost candidate technology to acquire digital form data. In this paper, we propose a pipeline to optimise, automate and accelerate the photogrammetric measurement process. The first step is to detect the optimum camera positions which maximise surface coverage and measurement quality, while minimising the total number of images required. This is achieved through a global optimisation approach using a genetic algorithm. In parallel to the view optimisation, a convolutional neural network (CNN) is trained on rendered images of the CAD data of the part to predict the pose of the object relative to the camera from a single image. Once trained, the CNN can be used to find the initial alignment between object and camera allowing full automation of the optimised measurement procedure. These techniques are verified on a sample part showing good coverage of the object and accurate pose estimation. The procedure presented in this work simplifies the measurement process and represents a step towards a fully automated measurement and inspection pipeline.
|