How Mississippi DOT Manages 1,300 Traffic Cameras with Wowza Streaming Engine

Mississippi DOT (MDOT) has been managing one of the larger statewide traffic camera video networks in the country for decades. If you’ve read the MDOT + Wowza case study, you already know the infrastructure story. MDOT treats Wowza Streaming Engine as a programmable, API-driven foundation for provisioning, monitoring, and delivering over 1,200 camera feeds across the state. That story is about how they built a platform that scales.

We recently sat down with Clint Johnson, Enterprise Systems Architect at MDOT, for a live webinar where he walked through the system his team has built, how they operate it day to day, and where they’re headed next with AI-powered video intelligence.

How To Make The Most Out Of Limited Monitoring Resources

In the operations center in Jackson, Mississippi, a small team of operators sits in front of video walls displaying 32 camera feeds at a time, cycling through tours of the entire statewide network. Each tour dwells on a set of traffic cameras for 30 seconds before rotating to the next batch.

During the day, the workload is distributed. Regional centers along the Gulf Coast and in Hattiesburg handle their local cameras, which reduces the volume that Jackson operators need to watch. But at night, it all consolidates to the overnight team in Jackson, which takes responsibility for every camera in the state.

With roughly 1,300 RTSP cameras in the public-facing network and 32 on screen at a time, it takes about 16 minutes to cycle through one complete tour. If an incident happens right after a set of cameras rotates off screen, it could be 15 minutes before an operator sees it. If a crash is blocking a lane during rush hour, or a vehicle is stalled on the shoulder in a construction zone, “That’s a 16-minute period of time before getting back to the camera that may have had a problem for the last 15 minutes,” Clint explained.

During that gap, operators may pick up information from other sources, like 911 feeds or reports from field crews. But the cameras themselves, the richest source of real-time visual information available, are only as useful as the number of eyes watching them. Any organization operating a large camera network faces the same constraint. Whether it’s a state DOT with a thousand highway cameras, a transit authority monitoring rail corridors, or a large venue managing security feeds across a campus, the bottleneck always boils down to humans only being able to watch so many screens at once.

How The Right Infrastructure Enables Video Intelligence

You can’t add intelligence to a video network that’s held together with manual processes and spreadsheets. The infrastructure has to be programmable.

Over the past several years, Clint and his team have built a custom management platform they call “Camera Manager.” It serves as the central system of record for every camera in MDOT’s inventory. This mission-critical platform leans heavily on the Wowza Streaming Engine API to automate what used to be manual work. Because all of it happens through the API, nobody needs to touch a Wowza configuration file to do things like:

• Adding a new camera
• Assigning a camera to a regional Wowza server
• Provisioning camera streams
• Migrating cameras between servers when hardware fails

MDOT uses the Wowza Transcoder to produce still image thumbnails from every active camera feed. By lazy-loading thumbnails from all 1,100 cameras onto a single page, MDOT can pull fresh images directly from Wowza servers distributed across the state. What’s more, these thumbnails load nearly instantly and are generated in near real time. This provides the exact foundational building blogs any AI vision model needs as input.

Another key element to MDOT’s infrastructure is ongoing system and hardware health checks. Every two hours, Camera Manager runs automated tests against the full camera inventory, checking response times, connectivity, and stream health. That systematic, API-driven monitoring is the same pattern used to feed frames into an AI pipeline at scale.

Understanding What’s Happening Within Video Streams

MDOT has been exploring computer vision for traffic cameras for several years. Clint’s team even experimented with traditional object detection models, the kind that draw bounding boxes around objects in a frame and label them as car, truck, pedestrian, fire truck, etc.

Pure object detection has a lot of valuable applications, but it also has a clear ceiling. It identifies objects, not situations. For example, it can easily tell you that there are vehicles in the scene. But, it cannot tell you as easily whether a crash has occurred, traffic is backing up behind a stalled vehicle, or if emergency responders are already on site.

The shift to vision models raises the bar for understanding and intelligence. Instead of drawing boxes, they interpret the full context of a scene. You can ask them a question about what’s happening in an image and get a structured, reliable answer back. They can not only track a car in a video stream, but also whether that car is moving or stalled, for example.

At MDOT, Clint and his team pull a still image from a camera via the Wowza Transcoder, then send it to a vision model with a specific prompt. MDOT already had the infrastructure to generate still images from every camera at speed. Because of that, the barrier to plugging in a vision model was significantly lower than it would have been. The prompt asks a series of yes-or-no questions, like:

• Has there been a crash?
• Is a vehicle stalled?
• Are there indications of a fire?
• Are any emergency vehicles present?
• Have the police arrived?
• Is a tow truck onsite?

The model returns a structured response for each question. “We can have the prompt feed us back a reliable response for each one of those scenarios,” Clint said. “If they’re all false, we move on to the next image.” If any flag is positive, that camera enters a priority queue for operator review. Per Clint, this is a fundamentally different capability than what was possible even a few years ago.

How MDOT Scans 1,300 Traffic Cameras Every Minute

One of the most impressive things about Clint’s system is that it scans all 1,300 cameras in MDOT’s network every minute. And it does so continuously, around the clock, without fatigue or shift changes. Compare that to the 16-minute manual cycle with MDOT’s staff, and the value speaks for itself.

Operators can now have a monitor on their desk displaying a priority queue. When the AI detects something noteworthy, it surfaces the alert with the triggering still image and a severity ranking. A crash ranks higher than a stalled vehicle. A multi-vehicle incident with emergency responders ranks higher still.

“They can see the still image that we captured that triggered that particular alert,” Clint explained. “They can click on another button and look at the live video to see what’s happening now, and then they’ll switch over to their ATMS system and make decisions based on that information.”

AI handles the scanning, but the operator stays fully in the loop. Humans still make important calls, coordinate responses, and communicate with the public. When an operator confirms an incident, they can push alerts to digital message signs miles upstream, giving drivers enough lead time to reroute. That’s the kind of response that prevents a slowdown from becoming a secondary crash. It also delays the need for costly lane expansions. Clint noted that, when he started at MDOT, the cost of adding a single lane mile was estimated at $5 million. Now, that number has almost certainly gone up.

Interested in Video Intelligence? Join The Beta!

Any operation that relies on human monitoring of large camera networks, from law enforcement surveillance systems to stadium and venue security, faces a version of the same scaling problem. The pattern is the same: generate frames, apply AI scene analysis, surface what matters, and let people make the decisions.

Wowza has been working on video intelligence capabilities that bring AI-powered analysis to live video streams. We’re currently in beta with several clients, exploring how to make this kind of scene understanding accessible to teams that already run Wowza Streaming Engine in their infrastructure.

The core idea is to take frames of video from existing streams, run them through inferencing services that extract actionable information, and deliver that information back to operators in a format they can act on quickly. MDOT’s work is a perfect example of what becomes possible when you pair a mature, API-driven streaming platform with modern AI. The infrastructure Clint’s team spent years building, the automated provisioning, the transcoder-driven thumbnail pipeline, the centralized system of record, turned out to be exactly the kind of foundation that AI needs to operate at scale.

“We went from a 15-minute delay potentially to maybe a one-minute delay,” Clint said, “and then putting out a message on a digital message sign five miles up the highway explaining to the public what to expect.” That’s the outcome that matters. Not AI for the sake of AI, but important information reaching the right people when they need it most.

Watch The Mississippi DOT Webinar On-Demand

If you’re exploring how AI-powered video intelligence could work within your operation, we’d like to hear from you. Wowza is still accepting participants into the beta program for organizations looking to bring scene analysis to their live video workflows. Build intelligent video capabilities into your technology stack, whether you are in the Transportation industry or not. Watch the on-demand webinar here, and reach out if you have any questions.