Host institution: University of Cambridge, United Kingdom
Project Title: Infrastructure objects detection based on cascaded deep learning architectures enhanced with design priors
Objectives: Conduct collaborative research on empowering top-down contextual detection strategies (e.g. detect building floors, then floor rooms, then room walls, etc.) with deep learning architectures and online learning up to the point of technological transfer to Trimble. The ESR must: (1) conceive the whole system; (2) define the critical design parameters; (3) implement and test three separate architectures into a common object detection prototype; (4) be involved in the subsequent development steps towards the realization of a commercial package.
Expected Results: The ESR is expected to realize and demonstrate an objects detection prototype able to detect most common objects of facilities and infrastructure (structural components, piping networks and mechanical components, doors/windows, etc.) when trained with a limited dataset and able to improve its detection performance with every additional test case, such that only uncommon, specialty items are left to be modelled by the user. The main bulk of the training will be at University of Cambridge, complemented by STFD, on deep learning computational geometry methods. Extensive training at LocLab Consulting on high-performance computing will also be indispensable. A final secondment to Trimble will explore the feasibility of product development.
Enrolment in Doctoral degree: University of Cambridge, PhD Engineering.
Advisor: Ioannis Brilakis