Alternative 3D processing techniques for complex automated material handling

Abstract

This project presents a new method for applying classical image processing and morphology to 3D vision systems. The motivation for this project is to allow for the further development of Autonomous Guided Vehicles (AGVs), specifically in their ability to interact with complex objects that cannot be characterised by current learning data sets, placed in un-structured environments. Current trends in machine vision research apply machine learning techniques to detect, isolate and analyse objects in complex environments. This research has shown how machine learning methods are proficient in the isolation of Objects Of Interest (OOIs), such as damaged and disorganised palletised carboard boxes, from environments that are cluttered with unwanted objects of no interest. An example of which includes palletised boxes, common household items and other objects that are repeatable or well documented in machine learning data sets such as Objects in Context. In addition to this, these systems are also capable of classifying the OOI within a narrow range of desired OOIs. However, the limitation of this approach to machine vision is the dependency on large data sets for the OOI in multiple positions and environments to train the machine learning systems. This is currently not a feasible approach for 3D machine vision, as relevant 3D data is not as readily available or as simple to capture as the conventional 2D image data used in the State Of the Art (SOA) machine learning techniques, via data sets such as Microsoft COCO (Common Objects in Context).

This thesis is embargoed for five years until 31/07/2030

Date of Award	Jul 2025
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	Combilift ULC
Supervisor	Joseph Butterfield (Supervisor), Adrian Murphy (Supervisor) & Karen Rafferty (Supervisor)