A research team at UNIST's Graduate School of Artificial Intelligence has developed a technology to improve the spatial recognition accuracy of camera-based autonomous driving systems by applying the concept of "vanishing point," utilized by Renaissance painters, to artificial intelligence.

A research team led by Professor Kyung-Don Joo of the UNIST Graduate School of Artificial Intelligence announced on the 15th that they have developed an AI model called "VPOcc" that compensates for perspective distortion in camera images. This research was led by first author Junsu Kim, and was jointly conducted by researcher Junhee Lee and researchers from Carnegie Mellon University in the United States.

Self-driving cars perceive their surroundings using cameras or LiDAR sensors. While cameras are cheaper and lighter than LiDAR, they often suffer from distance distortion when representing three-dimensional space as two-dimensional images, leading to the loss of distant objects or the emphasis on nearby areas.

The research team solved this problem by applying the "vanishing point," a concept used by Renaissance painters to express perspective, to artificial intelligence. Vanishing points are points where parallel lines converge in the distance, and are used by humans to perceive depth on a flat surface. VPOcc reconstructs image information based on these vanishing points, more accurately restoring depth and distance.

The model consists of three modules. VPZoomer reduces perspective distortion in an image based on a vanishing point, while VPCA extracts balanced information from near and far areas. SVF combines the original and corrected images to compensate for each other's weaknesses.

Experimental results showed that VPOcc outperformed existing models in terms of spatial understanding (mIoU) and reconstruction accuracy (IoU). In particular, it successfully predicted distant objects more clearly and distinguished overlapping objects more accurately in road environments, which are crucial for autonomous driving.

Researcher Kim Jun-su said, “We wanted to improve understanding of 3D space by incorporating the human spatial perception method into artificial intelligence,” adding, “This will be a great advantage in terms of price competitiveness and lightweighting of camera sensors.”

Professor Joo Kyung-don said, “This technology can be applied to various fields, including autonomous driving, robotics, and augmented reality (AR) map production.”

This research was supported by the Ministry of Science and ICT and the National Research Foundation of Korea, and its findings were accepted for publication at IROS 2025, an international conference on intelligent robotics. This year's conference will be held in Hangzhou, China, starting October 19th.