Research by Rozenn Dayhot Vladimir Krylov to assist in the mapping of physical assets by Ireland’s national telecommunications operator (EIR).
In their work, the team developed a hybrid approach for using Machine Learning exploiting (Google) Street View with dynamic object detection and Mapping. The approach was used in an Industry Collaboration project with EIR
The available coverage of the earth’s surface in ariel and street-level imagery has grown dramatically in recent years. Street view imagery openly available from Google, Bing, and Mapillary represent a collection of over 100 billion geo-tagged images covering millions of kilometers of roads worldwide. In Ireland alone since its launch in 2010, Google Street View has covered over half of the country’s public road network, with most areas in urban centres captured multiple times. This high-resolution imaging modality constitutes an invaluable source of information about the road network environment that could benefit many applications including autonomous navigation, urban planning, and asset management.
In this Eir project, we designed an innovative software solution that performs automatic detection and geo-referencing of objects from street view images. Types of objects include street furniture (electrical, telephone, postal, road, and state-owned infrastructure), facade elements (house numbers, shop signs, cameras, antennas, security alarm boxes), minor landmarks, etc. The solution is customisable to detect a specific type of object; in eir’s case, telegraph poles. The solution is based on a fully automatic algorithm that consists of state-of-the-art deep learning elements for image analysis and an innovative geolocation module. The experimental analysis demonstrates an over 90% success rate for object detection and geotagging accuracies comparable with those of commercial-level GPS units used by human operators. The project team comprised of four ADAPT researchers and six EIR representatives.
Inventory of road furniture and the road environment needs to be routinely monitored to ensure the safety of road users: damaged assets need to be replaced; occluding vegetation needs to be cleared to improve visibility; potentially dangerous obstacles need to be removed. The ADAPT Research Centre software pipeline provides a solution that can assist in automated road asset management and monitoring. The unique and innovative software solution can be integrated with existing asset management solutions to complement asset inventory through automatic image analysis. The solution facilitates asset monitoring by identifying elements in need of maintenance, reducing the level of human involvement. Other service industries such as logistics, monitoring, and security, road authorities as well as architects and designers of the built environment can avail of the ability to automatically and remotely detect artefacts from any location where street view imagery is available.
“ADAPT’s research team helped us by using sophisticated image processing and machine learning to get up-to-date information about our pole inventory. The project will have significant operational and planning opportunities for eir and other operators.” Helena Graham, CTO
This is the first, and to our knowledge, the only automatic tool that jointly addresses detection and geo-location of objects from street view imagery. The innovative solution can be customised to detect any specific stationary objects that can be found in the vicinity of the road network. The existing asset management solutions are based on manual data analysis or target special cases that do not generalise to other object classes (e.g., street names detection system by Google). The software solution is scalable in terms of both geographic extent and classes of objects to be detected. The deployment could be used at a county or national level for one or more types of objects that need to be mapped. The imagery may be harvested automatically from providers like Google, Bing, and Mapillary, or collected by custom imaging systems designed for specific applications.
The technology has the potential for other applications across a range of fields including road maintenance; agricultural and environmental monitoring; utilities; logistics; and as an enabler for the Geospatial Semantic Web.
