Lab Manager | Run Your Lab Like a Business
Self-driving car on the highway
iStock, 3alexd

New 3D Object Detection System Promises Improved Autonomous Vehicle Safety

Researchers develop an object detection system that uses deep learning and IoT for fast, local obstacle navigation

by
Holden Galusha

Holden Galusha is the associate editor for Lab Manager. He was a freelance contributing writer for Lab Manager before being invited to join the team full-time. Previously, he was the...

ViewFull Profile.
Learn about ourEditorial Policies.
Register for free to listen to this article
Listen with Speechify
0:00
5:00

Researchers from Incheon National University in Korea have developed a new 3D object detection system specialized for autonomous driving. The system uses deep learning and IoT for fast, on-site data processing and efficient navigation of various obstacles.

"For autonomous vehicles, environment perception is critical to answer a core question, 'What is around me?'" explained professor Gwanggil Jeon, who led the study. "It is essential that an autonomous vehicle can effectively and accurately understand its surrounding conditions and environments in order to perform a responsive action."

The system is based on You Only Live Once, Version 3 (YOLOv3), a well-known identification algorithm that was first used for 2D object detection and then modified for 3D objects. The team fed RGB images and 3D measurement points, also called point cloud data, as input to YOLOv3. The algorithm returned classification labels (recognizing what an object is) and bounding boxes (rectangular box that contains the recognized object) with confidence scores quantifying how sure it was of each recognition. They then tested its performance with Lyft’s Level 5 Open Dataset. The early results revealed that YOLOv3 achieved over 96 percent accuracy for 2D and 3D objects, outperforming competing detection models.

"Our system has the potential to be integrated into a wide range of applications where object and obstacle detection, tracking, and visual localization is required," Jeon said. According to Jeon, current self-driving vehicles rely on technology like Light Detection and Ranging (LiDaR), but it is predicted that LiDaR will be replaced with general cameras and deep learning algorithms. Jeon believes his team is at the forefront of this shift.

The team's findings were recently published in the IEEE Transactions of Intelligent Transport Systems journal. Jeon is optimistic about the future of autonomous vehicles, saying, "Based on the development of element technologies, autonomous vehicles with improved safety should be available in the next five to 10 years."