A technology capable of reducing the precision correction limitations of event cameras, which are robust against fast movements and low-light environments, has been developed by UNIST. The research team proposed a method to locate checkerboard reference points without converting event data into standard images, and the paper was selected as a CVPR Highlight Paper in recognition of its potential for application in high-speed vision systems, including robotics, autonomous driving, and AR/VR.

Professor Kyung-Don Joo's team at the UNIST Graduate School of Artificial Intelligence announced on the 1st that they have developed a computer vision calibration technology that detects reference points of event cameras using a checkerboard calibration plate. This research was co-authored by UNIST researcher Tae-Hoon Ryu as the first author and researcher Chang-Woo Kang.

Camera calibration is the process of reducing perception errors caused by lens distortion. In standard cameras, the degree of distortion is calculated based on the vertices of a checkerboard grid captured from various angles.

The event camera does not save the entire screen frame by frame, but records only the point where a brightness change occurs as an event. While it is advantageous for fast movements, the existing method was difficult to apply because events rarely occur at the vertices of the checkerboard.

The research team first analyzed grid lines where events were relatively distinct, and then calculated the point with the fewest events in the area where the lines intersect as a reference point. They utilized the characteristic that brightness changes at the vertices cancel each other out, resulting in less event information.

The problem of grid lines becoming blurry due to movement was also addressed. The research team made the grid lines sharper by aligning events occurring at different points in time to a single reference point.

This technology was also applied to AprilTag detection, which is used for location recognition in robots or AR/VR devices. The research team identified the shape and number of the markers using only event data and detected markers that were visible even when partially obscured.

Researcher Ryu Tae-hoon explained, “This technology can improve correction accuracy by finding a reference point within the signal recorded by the event camera itself.”

Professor Joo Kyung-don stated, “Accurate camera calibration is the starting point for various vision technologies,” adding, “I expect it to serve as a foundation for expansion into robots, autonomous driving, and AR/VR systems.”

This research has been selected as a highlight paper for CVPR, an international conference in the field of computer vision, which will be held in Denver, USA, for five days starting on the 3rd. Highlight papers account for only about 3.5% of all submitted papers.