Using Deep Learning for Roads and Infrastructure

We were recently invited to present at the Transport Technology Forum, where we spoke about how we are working in collaboration with Croydon Council on a pilot project using Deep Learning funded by the Department for Transport,  for the ‘Connected Vehicle Data Competition’.

In the context of the £5bn regeneration programme in Croydon, the challenge Connected Space was presented with and the aim of the pilot is to mitigate impacts of construction traffic on Croydon’s roads, by looking at how construction vehicles - especially HGVs - use road assets, as well as assessing any current damage, and ultimately fixing road condition issues.

Deep Learning Driven Solution

The approach we are taking to solve the challenge given is to collect road image and video data from vehicles driving along the construction routes, and using deep learning, analyse the data and identify road asset issues such as potholes. Once analysed, this data can then be fed back into our data platform, known as the Urban Data Platform, where council officers or contractors can see and receive notifications of identified issues.

The Urban Data Platform can be used as a reporting tool, and provides the ability to overlay multiple other data sources - such as Construction, Traffic, Roadworks, Air Quality, and now Road Quality - to provide visibility over the specified area.

Pilot Progress

Applying an iterative and agile approach, we have been collecting data from a range of sources over the course of the project. We started out by taking images near our studio in Croydon, and even used YouTube videos to train the model. We are now collecting data from the dashcams on vehicles driving around on the construction routes and can use the video data from an asset management fleet in Croydon.

The collected data can then feed into the Urban Data Platform integration we have built, and there is an API to feed into other third party systems if required.

We have also enabled the model to be performant on mobile. This means it can perform (as in speed) near real time - enabling greater reach and a more simple, convenient, and cost effective solution, which will open up wider opportunities, and more application use cases.

What Next?

Great progress has been made and the results are very promising! However, we need to continue to develop and trial the solution in order to deliver on the outcomes, improve the the model, and gather further data to develop the business case.

To achieve this, we are looking for partners to work with including Local Authorities (to share findings with, and any that are interested in discussing a pilot or trial), Fleet Operators (who can help gather more data), and Commercial Partners (to help develop the business model and monetise the innovation).

For more information on the project and how Deep Learning can be used in your organisation, get in touch with David Gregory here