This AI Catches Trouble by Watching How People Interact
Security cameras could soon spot danger that looks perfectly normal to us.
Most camera AI works like a strict hall monitor: it looks at each person and each object separately and asks, "Does this look normal?" That approach has a blind spot. Two people standing near a door are fine. One person standing near a door is fine. But one person slowly following another, matching their every turn, is not — and the old systems often miss it, because nothing looks strange on its own. Researchers from China have now published a system that watches the relationships instead.
The framework is called INTERACT, and it has three moving parts. The first studies every person and object in a frame together — how they look, how they move, and where they sit relative to each other. The second uses that history to predict what should happen next: if a hand usually stays away from a bag, it notices when it drifts closer. The third double-checks those predictions for meaning, not just error, so the system catches oddities that a simple "that's not what I expected" rule would wave through. Crucially, it never needs to know anyone's identity.
On ShanghaiTech, a widely used research video dataset, INTERACT scored 84.5% — a new high for this kind of unlabeled training. When the researchers removed the part that links entities together, performance dropped the most, proving the relationship idea is doing the heavy lifting. The honest catch: this is a lab benchmark, not a real street. Real cameras face bad weather, crowds and weird lighting, and any anomaly detector will cry wolf sometimes.
The bigger picture is practical. Patrol robots and city cameras are getting cheaper and more common, and this kind of software could make them genuinely useful rather than just recording. It could flag a fall, a fight or a theft in seconds instead of after the fact. It also sharpens an old trade-off: the same tool that protects you is the one watching you. The code is promised to be public, so expect this capability to spread quickly.
- Older camera AI judges people and objects one at a time, so it misses suspicious situations where everyone looks normal individually
- INTERACT checks relationships instead — who is near whom, who is following whom — and scored 84.5% on a standard video test
- The team will publish the code for free, meaning surveillance systems and patrol robots could pick it up quickly
Why It Matters
Security cameras and robots may soon flag real danger in seconds — while watching all of us more closely.