The South Africa National Roads Agency Limited's (Sanral's) Technical Innovation Hub (TIH) is exploring the use of machine learning to improve road safety, reduce congestion and inform infrastructure development.
Information provided by machine learning can be used to activate the appropriate response through the Road Incident Management System (RIMS), remedy the situation and inform road users in real time, explains TIH mechatronic engineer Ruan van Breda.
“Machine learning can be used to detect and segment and classify objects within a camera frame based on pre-trained image classifiers. Within the road environment, this allows one to detect and classify different types of vehicles, pedestrians, animals or cyclists, among others. There is already ample data available to classify objects according to these types.”
The classification classes can be further expanded through the creation of custom data sets and training classifiers to be able to distinguish, for example, between slow moving traffic and a road traffic crash.
This can also be used to create new classification classes based on unique experiences or the requirements of the road authority, such as fire or protest detection, foreign objects such as rocks, or tyre detection, besides others, he points out.
“One can also look at how these different objects interact with one another to detect unusual vehicle behaviour, like a vehicle stopping on the freeway. One can also infer information about the interaction between multiple elements such as cars and pedestrians. If a vehicle is detected moving to the side of the road and coming to a standstill and pedestrians are detected moving towards the vehicle and enter the vehicle, this can be classified as an informal pick-up,” Van Breda explains.
As more data is collected, these trends can either aid road authorities with infrastructure planning such as drop-off or pick-up points, or aid law enforcement to stop illegal pick-ups, if it is considered a safety risk, he adds.
“While this technology is still in the exploratory phase in South Africa, countries like China use machine learning to incorporate facial recognition for law enforcement. They are able to identify the driver of a vehicle and instantly issue fines, if that driver does not have a valid driver’s licence. Fines can also be issued automatically for individuals who jay-walk or gain access to restricted areas.”
Technology of this nature also comes with significant risks and all efforts are being made to understand how to effectively use the technology while maintaining strict compliance with legislation as it relates to the privacy of road users.
“Some of the ways to mitigate these potential privacy risks are to use strict security and access controls. Data can also be anonymous at the point of capture. After all, the intention is not to observe individuals, but rather to identify trends and incidents to inform appropriate response and interventions,” Van Breda says.