Lidar and AI combine to help improve road safety

This approach focused on the most vulnerable road users, namely pedestrians and cyclists, and could be used to help improve driver awareness, reduce accidents and better understand behaviour in work zones.

In a recent study, a team led by associate professor Yaw Adu-Gyamfi and graduate student Linlin Zhang at Missouri’s College of Engineering created a new method to understand how pedestrians, cyclists and vehicles interact, particularly at traffic signals.

“By having a better understanding of how pedestrians and cyclists interact with each other on the roads, this study will help us design advanced systems that will allow vehicles to better understand and avoid other road users. This is important especially as autonomous vehicles become more common,” Adu-Gyamfi said in a statement.

The technology can help spot near misses between cars and pedestrians, allowing experts to better understand how to prevent accidents. As it becomes more widely available, it could track how people and cars approach intersections and share that data with vehicles.

“This approach would require working with car manufacturers to build the technology into vehicles,” said Adu-Gyamfi. “In fact, some cars already connect with traffic systems using networks like cellular vehicle-to-everything (C-V2X).”

The data collected by this system could be used in other ways to improve transportation such as helping specialists decide how long pedestrians need to cross roads safely. It could also track cars entering work zones, catching speeding or distracted drivers. Plus, it can spot pavement problems, such as the depth of potholes.

For this project, the researchers set up a joint camera and lidar system at an intersection to monitor traffic flow. Instead of the traditional approach that requires using two lidar units, they optimised the technology to work with one unit. Also, by applying point cloud completion, they were able to improve the visibility of pedestrians and other objects over existing methods.

“Instead of retraining a machine learning model to detect objects, we used a pre-trained one and created a new algorithm to estimate an object's height and width,” said Adu-Gyamfi. “This helped us classify objects, such as buses, pedestrians and cyclists, more accurately than other AI models designed for the same task.”

Before this technology can be widely used on roads and highways, researchers will need to address challenges with data processing, power supply stability and weather conditions.

The study, “Three-Dimensional Object Detection and High-Resolution Traffic Parameter Extraction Using Low-Resolution LiDAR Data,” is published in the Journal of Transportation Engineering. Co-authors are Xiang Yu at Mizzou and Armstrong Aboah at North Dakota State University.