Abstract
Point clouds are unordered sets of points that can be used to represent 3D objects. Using point clouds to identify defects in manufactured parts is an area that is still being researched. This research aims to identify artificial intelligence architectures that can effectively work with point clouds for defect detection in aerospace manufacturing. Background segmentation was selected as a representative task for evaluating these architectures. Three model architectures were analyzed (a fully connected neural network, a convolutional neural network, and PointNet). The architectures were tested on the ScanObjectNN dataset containing noisy point clouds. Out of all the methods tested, the PointNet model performed the best, working well with point cloud data and extracting meaningful information. Further research is required to identify a more optimal dataset for the use case, as well as evaluating other model architectures that can be used with point clouds.