Skip Ribbon Commands
Skip to main content

Helping Autonomous Vehicles Go Off-Road

Off-road environments can be challenging for autonomous vehicles when there are heavy shadows, steep inclines or muddy water puddles.

Researchers led by Professor Justin Dauwels at Nanyang Technological University’s School of Electrical and Electronic Engineering (EEE), Centre for System Intelligence and Efficiency (EXQUISITUS) have created a two-stage semantic segmentation system to help autonomous vehicles “see” off-road routes more clearly.

The system can not only differentiate between dirt roads and their surroundings, such as vegetation and sky, but also identify water puddles, which can look like dirt roads.

The invention will help autonomous vehicles to navigate off-road routes more easily and safely. An unmanned ground vehicle without the EEE system may get stick in water puddles on the road, for example, requiring costly human intervention such as towing.

To “see” the off-road route and identify water puddles, the EEE system relies on images from cameras and data from a 3D Lidar (light detection and ranging) rig.

It first uses the information to classify the environment in front of the autonomous vehicle into road and non-road segments. This helps the car to “see” the dirt road.

It then uses a specialised water puddle classifier, trained by the EEE researchers, to analyse the road segments and distinguish between water puddles and the road. This allows the car to “see” the puddles on the road and avoid them.

Dividing the process into two stages reduces the amount of work for the specialised classifier, since it need only examine the dirt road and identify water puddles on it. This improves its speed and accuracy.

The EEE researchers tested their system on eight videos containing 1,023 annotated frames. It achieved an F1 score of about 93 per cent for road detection and 80 per cent for water puddle segmentation in more than 10 hertz.

Professor Dauwels added that their system can be used in urban environments as well, for example to enable autonomous vehicles to distinguish lane markers from road regions.

By Professor Justin Dauwels

Click here to find out more.


Published on: 2-November-2017 ​​​​​

Not sure which programme to go for? Use our programme finder
Loading header/footer ...