Plane-based Object Categorization using Relational Learning

We use Inductive Logic Programming (ILP) to learn classifiers for generic object recognition from point clouds, as generated by 3D cameras, such as the Kinect. Each point cloud is segmented into planar surfaces. Each subset of planes that represents an object is labelled and predicates describing those planes and their relationships are used for learning. Our claim is that a relational description for classes of 3D objects can be built for robust object categorization in real robotic application. To test the hypothesis, labelled sets of planes from 3D point clouds gathered during the RoboCup Rescue Robot competition are used as positive and negative examples for an ILP system. The robustness of the result is evaluated by 10-fold cross validation. The results show that ILP can be successfully applied to recognise objects encountered by a robot in an urban search and rescue environment. Keywordsobject classification, inductive logic programming, ALEPH, cross validation, urban search and rescue, point cloud, machine learning, range data.