I spent this year’s ITS America in Detroit walking the expo floor and talking with different DOTs and traffic-management teams. One pattern stood out.
The established vendors are genuinely good at what they do: traffic counting, roadway inspection, over-speed enforcement. The catch is that most of them do only those things. Each product handles a fixed, narrow list of use cases, and if the problem you’re facing isn’t on that list, you’re waiting, sometimes months or even longer, for someone to build a detector for it.
The problem: real incidents don’t follow a script
The trouble is that incidents on the road don’t follow anyone’s list.
This is clearest in tunnels. A tunnel is an enclosed space that’s hard to evacuate, where every second of response time counts. Think about what can go wrong: a power outage that plunges a lane into darkness, a sudden loss of visibility from smoke or haze, a fixture falling from the ceiling onto the roadway, flooding, a wrong-way driver, a vehicle stopped in a live lane. Each one is rare. Each one is also a potential emergency. And a traditional, rule-based system that was never explicitly programmed to look for them simply won’t see them.
Operators know this. They can’t watch every camera at once, and the events that matter most are often the ones no one anticipated.
Traditional analytics and VLMs: better together
I want to be clear about something, because it’s easy to misread. We’re not here to throw out traditional video analytics. We use it and we believe in it — for the well-defined, high-volume tasks it does reliably, it’s the right tool.
What we add on top is Vision Language Model–based (VLM) detection, guided by our cynapse.ai ontology: essentially a structured understanding of what vehicles, road, lanes, people and events mean in a roadway or tunnel context. Pairing them lets us cover far more ground: faster to stand up, broader in what it can catch, and at the level of quality this kind of environment demands. Traditional analytics for the known and repetitive; VLM-based detection for everything else.
Automated incident detection and user-defined events
We call this Automated Incident Detection (AID). The part operators respond to most is what we call user-defined events.
Instead of waiting for a vendor to build a bespoke detector, you describe the situation in plain language such as, “alert me if a vehicle stops in a tunnel lane,” “flag any loss of visibility,” “tell me if something falls onto the roadway”, and the system starts watching for it.
A new scenario you’re worried about today can be a live detection tomorrow. That flexibility is the difference between a tool that fits your site and one your site has to work around.
From alert to Automatic Incident Report — in seconds
Detection is only half the job. In a control room, an alert that just says “something happened” still leaves an operator scrubbing footage to work out what, where and how serious it is.
So, AID doesn’t stop at the alert. cynapse.ai AI agents assemble the Automatic Incident Report for you including what happened, when, the clip, and the surrounding context, all in seconds rather than the hour it might otherwise take. The operator opens a clear picture instead of starting an investigation from scratch.
Why this matters for operators
Put together, that means fewer missed events, faster and better-informed response, and far less manual work for teams that are already stretched.
Where this is heading
The vendors that do one thing well will keep doing that one thing well. But what DOTs and agencies actually asked me for at ITS was different: a way to catch the incidents that don’t fit the template, and to be able to act on them faster.
And that’s the real shift: from fixed, pre-programmed detection to agentic video intelligence that understands context and helps you respond. It isn’t a someday idea. It’s here now.
About the Author
Senior Product Manager
Sinan Asil drives product strategy for cynapse.ai’s Agentic Video Intelligence platform, with a focus on cynapse Copilot and the product roadmaps for cyRoad and cyPort. He brings 15 years of experience building and scaling AI, analytics, and enterprise SaaS products.
