3D payload detection from 2D range scans

Payload recognition is an important ability for autonomous industrial vehicles. While scanning laser rangefinders are commonly used on autonomous vehicles, most provide only 2D range scans. The goal of this research is to use the 2D range data to accurately identify an initially unknown 3D asymmetric payload. Typical approaches to payload recognition use either artificial markers or models with relatively tight constraints on the shape. We relax these constraints and allow the system to determine the most appropriate representation of the target object using an unsupervised learning approach. A set of target scan segments from the training set is reduced to a reference set with a high discrimination capability. The system uses the reference set in a classifier that evaluates incoming scans and monitors areas in the environment that potentially contain target objects. Upon a high enough confidence, a target is declared. Once trained, the system is able to accurately recognise a target object in different environments. The reduced reference set classifier shows faster convergence to a target classification than one developed with a full feature set and another with k-means clustering.