ETRI sweeps top two spots at international traffic AI competition

- Takes first in pedestrian intent analysis, second in traffic violation detection at ECCV 2026 AI City Challenge

- System integrates CCTV, fisheye lens and vehicle camera footage, and explains the reasoning behind hazard assessments

Kim Sang-won, a researcher at ETRI's AI Infrastructure Research Laboratory, demonstrates the traffic video analysis technology. [ETRI]

A domestically developed AI system that detects the moment a pedestrian is about to cross a road — and autonomously identifies dangerous situations such as signal violations and wrong-way driving — has been recognized as world-leading at an international competition.

The Electronics and Telecommunications Research Institute (ETRI) announced Tuesday that a joint research team it formed with the University of Washington took first and second place overall in two separate categories at the 10th AI City Challenge, held Sept. 8 in Malmö, Sweden, as part of ECCV 2026, a leading international conference on computer vision.

The AI City Challenge is an international AI competition drawing researchers from global universities, research institutions and companies, including Nvidia. This year's event focused on evaluating whether AI systems can operate reliably in environments they have not previously encountered, and how accurately they can interpret and reason across footage from different cameras and angles.

The joint ETRI-University of Washington team, competing under the name UWIPL_ETRI, placed first overall among seven teams in the PSI-VQA category, which involves analyzing pedestrian crossing intent, and second overall among eight teams in the FETV category, which focuses on detecting traffic violations at intersections.

PSI-VQA requires analyzing footage from a vehicle's forward-facing camera to determine whether a pedestrian intends to cross the road, identify when a dangerous situation may occur, and explain the reasoning behind those assessments. The task goes beyond basic object recognition — detecting people or vehicles — to inferring behavioral intent.

FETV involves identifying violations such as running red lights, wrong-way driving and jaywalking from fisheye lens camera footage installed at intersections, and generating descriptions of each incident. Fisheye lenses can capture a wide intersection in a single frame, but image distortion increases toward the edges, making it difficult to accurately track the positions and movements of vehicles and pedestrians.

At the heart of the system is UniTraffic, an integrated traffic video understanding technology.

ETRI took first and second place in two categories at the 10th AI City Challenge. From left: Byeon Woo-jin, head of ETRI's Daegyeong Regional Research Division; Kim Gwang-ju, division director; researcher Kim Sang-won; and senior researcher Kim Byeong-geun. [ETRI]

Where conventional video analysis AI systems require separate models for each type of camera — such as CCTV or vehicle-mounted cameras — UniTraffic processes footage from ceiling-mounted CCTV, intersection fisheye lenses and onboard vehicle cameras through a single vision-language model (VLM)-based system.

A VLM is an AI that understands and reasons across both visual and linguistic information. Rather than simply recognizing people and vehicles in footage, it can describe in language what is happening and explain why a situation is dangerous.

UniTraffic first rapidly scans the full video, then focuses detailed reanalysis on scenes requiring precise judgment — such as pedestrian movements or vehicles changing course. The system is designed to process long stretches of traffic footage efficiently without missing critical hazardous situations.

The team also applied a Traffic Evidence Graph to improve the reliability of the AI's assessments. The technology links a pedestrian's location and direction of movement, the behavior of nearby vehicles, and the timing of events to build a basis for each judgment. This reduces the AI's tendency to generate content not present in the footage as if it were fact, and improves the explainability of its decisions.

"This is the result of advancing the intelligent traffic control and video analysis technologies we have built up over many years into cutting-edge AI through international collaborative research," said Byeon Woo-jin, head of ETRI's Daegyeong Regional Research Division. "We will expand real-world validation at actual traffic sites and deepen cooperation with domestic companies to develop this into traffic and safety services that people can experience firsthand."

nbgkoo@heraldcorp.com