NSW Transport and Microsoft use machine learning and data to reduce road accidents

The pair ran a trial to identify five potentially risky intersections.
Written by Aimee Chanthadavong, Contributor

Transport for New South Wales and Microsoft have partnered to develop a proof of concept that uses data and machine learning to flag potentially dangerous intersections and reduce road accidents.

As part of the proof of concept, Transport for NSW ran a trial in Wollongong to uncover five potentially risky intersections. It involved 50 vehicles generating more than a billion rows of data over a 10-month period, before Databricks and Azure were used to curate, ingest, and interpret the data.

The telematics data was used to identify speed, harsh braking, harsh acceleration, and lateral movement just before the intersection. It was then compared to patterns of existing crash investigation data.

"We had a circle of interest around the intersection … when the vehicle is actually approaching the intersection, how does it behave at 50 metres, how does it behave at 25 metres, and how does it behave going through the intersection?" said Julianna Bodzan, head of data discovery program at Transport for NSW.

While the five intersections were not previously known as dangerous spots, the data, according to Microsoft, based on analysis of driving patterns revealed otherwise.

Since the trial, two out of the five intersections have been scheduled for modification.

Bodzan said identifying dangerous intersections on roads was just one area where Transport for NSW would look to use AI and machine learning.

"Transport is not just roads. We look at railway lines, we are thinking about pedestrians and school zones -- and we are thinking about how we can use this type of data and enrich with other types of data to basically improve the safety of the network," she said.

The work being carried out by Transport for NSW aligns with the Centre for Road Safety's strategic plan to achieve zero fatalities and serious injuries on state roads by 2056.  

Related Coverage

Editorial standards