Turning Existing Traffic Cameras Into Safety Sensors

A pedestrian is killed by a vehicle roughly every two minutes somewhere in the world. Behind almost every one of those collisions is a pattern that was visible beforehand. Vehicles taking a prohibited turn at the same corner. People crossing late against the signal. Speeds well above the posted limit at a particular hour. The warning signs were there. Nobody was counting them.

The Cameras Are Already There

Municipalities have spent years installing cameras at major intersections. Those cameras record continuously. Almost all of that footage is never watched. It is kept for a period, used to reconstruct an incident if one happens, and then overwritten.

That is an extraordinary amount of information about how an intersection actually behaves, sitting unused. Computer vision changes what it can be. Instead of a recording, each camera becomes a sensor that counts, classifies, and flags, continuously, without a person watching a screen.

What Becomes Measurable

Once a system is watching, a great deal becomes countable that used to require a manual traffic study. Vehicle volume by approach direction. Movement classification. Heavy vehicle share. Headway between vehicles. Stopped time and Level of Service. Optical speed estimation against the posted limit, including the 85th percentile speed engineers rely on.

The same system can track people. Curb to curb crossings by crosswalk and direction. Crossings outside the crosswalk. Late crossings measured against the WALK phase. Average crossing speed. Cyclists and scooters counted separately, in dedicated lanes or mixed traffic. Live signal state for both vehicles and pedestrians, so every event is understood in the context of what the lights were showing.

From Counting To Preventing

Counting is useful. The safety value comes from conflicts. A near miss between a pedestrian and a turning vehicle is not an accident, but it is the closest thing to a rehearsal for one. When a system logs those conflicts by type, pedestrian and vehicle, cyclist and vehicle, vehicle and vehicle, a pattern appears long before anyone is hurt.

That pattern is what planners need. Instead of waiting for collision statistics, which take years to accumulate and arrive too late, a traffic engineer can see this month that a particular approach is generating conflicts and adjust signal timing, signage, or geometry now.

And when a collision does happen, real time incident detection means first responders are notified as it occurs rather than when someone finds a phone.

Doing It Responsibly

Watching public space demands care. Face blurring and license plate blurring should be available options, so the system counts behaviour without identifying people. Accuracy should improve over time through classifier review and retraining, with the municipality in control of that process. Reports should export in open formats, PDF for reading and CSV or JSON for analysis, so the data belongs to the city, not the vendor.

No New Infrastructure

The most important point is the simplest. None of this requires radar, new poles, or new cameras. It requires software that can calibrate to any intersection layout and read the feeds that already exist. That is the idea behind our Traffic Vision product, and it is why a municipality can start with a handful of intersections rather than a capital project.

The footage has always been there. The difference is whether anything is learning from it.