Enhancing Traffic Safety: IREX's Next-Gen Pedestrian Detection

for demo purposes only

MURRIETA, CA - IREX, a global leader in Ethical AI and intelligent monitoring based in the U.S., is updating its automatic pedestrian crossing violation recording with a new focus on more accurate pedestrian detection in real-world traffic scenarios. At busy crossings, even a small error can affect whether a driver is detected violating traffic laws. The update is focused on one of the most difficult tasks in traffic analytics: reliably identifying when a driver fails to yield to a pedestrian at a crosswalk.

The IREX system uses two parallel detection streams: one tracks pedestrians in the crossing zone, while the other tracks vehicles in the slowdown zone. If a vehicle leaves the slowdown zone while a pedestrian is still in the crossing and moving toward that vehicle, the system records a violation. This sounds straightforward, but one detail makes all the difference: the system must understand the pedestrian’s direction of movement and the relationship between the pedestrian and the vehicle. If a person is walking away from the vehicle, or has already left the crossing area, no violation should be recorded. That’s why accurate movement analysis is just as important as object detection itself.

Built for realistic crosswalk conditions

In practice, crosswalk footage is far more complex than controlled training data. Top-down CCTV cameras capture zebra crossings, traffic lights, signs, road markings, passing cars and dense traffic. In this environment, standard perimeter-focused models can make mistakes. In busy urban scenes, these models may confuse distant objects, partially visible figures, or even passengers inside vehicles with pedestrians. These errors affect tracking stability and can lead to false positives, resulting in false alarms, while in other cases, pedestrians may be missed altogether.

for demo purposes only

To resolve this, IREX is developing a new AI-training pipeline based on actual crosswalk footage, with data that better represents the target environment. This process involves filtering current perimeter data and incorporating new, domain-specific material.

Next steps in development

Two model paths are being tested: fine-tuning the existing perimeter model and training a new model from scratch. The IREX team is currently working with a set of real-life videos, labelled manually and prepared for validation and testing.

For now, the focus remains on pedestrians only. Detection of personal mobility vehicle users is planned for a later stage. Once detector performance is reliably improved, the tracking component will be reviewed as well. If the current tracking quality is not sufficient, IREX will move to more advanced tracking methods based on neural networks and descriptors to meet the required quality standards.

for demo purposes only

“Our priority is simple: to help our models read a crosswalk the way a trained operator does,” said Serge Smirnoff, spokesperson for IREX. “When pedestrian detection improves, everything downstream improves with it — fewer false alarms, steadier tracking, and violation records we can actually stand behind in real traffic.”