New AI Predicts When Pedestrians Will Step Into Traffic
Self-driving cars that read body language could mean fewer accidents and safer streets.
Researchers François G. Landry and Moulay A. Akhloufi have published a new AI system called TrajFusionNet+ (say that five times fast). Its job is simple to describe and hard to do: look at a pedestrian on a street and predict whether they're about to walk into traffic. That's the exact question a self-driving car needs answered a second or two before it matters. If the car knows someone is likely to step out, it can slow down early instead of braking hard at the last moment.
What makes this approach different is that it doesn't rely on one clue. It combines three ways of looking at the scene. First, it tracks how a person has been moving — speed, direction, whether they've been drifting toward the curb. Second, it looks at the actual camera picture, drawing boxes around where the person is and where they're likely to go. Third, it maps out the relationships in the scene: is there a crosswalk? A parked van blocking the view? Another pedestrian already stepping off the curb? People often cross because someone else did, and this model tries to capture that.
On two widely used research datasets — called PIE and JAAD, both built from real driving footage — the new model outperformed previous approaches. The authors also tested something new: training the AI on both datasets at once but testing it on each separately. That's a check for whether the AI actually learned general rules about human behavior, or just memorized one particular set of streets. Their model held up better, which suggests it might travel well to new cities and new kinds of roads.
The honest catch: this is a research paper, not a product. It's been submitted to a journal and hasn't been through full peer review. The datasets are limited, filmed in specific places with specific cameras, and real streets are messier — bad weather, crowds, people on phones, cyclists weaving through. Nothing here means your car can read minds today. But each improvement like this is a small step toward cars that anticipate people, rather than just react to them.
- The AI predicts whether a pedestrian will cross by combining how they're walking, what the camera sees, and how they relate to nearby traffic.
- It beat previous methods on PIE and JAAD, the two standard test sets for this problem.
- It's early-stage research — not in any car you can buy, and still untested in messy real-world conditions.
Why It Matters
Safer self-driving cars mean fewer pedestrian deaths — and less nerve-wracking moments for everyone sharing the road.