Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / How Edge Computing Is Rerouting Traffic in Smart Cities
Transcript
- Lucas: There's this number that's stuck in my head — the average American commuter loses 54 hours a year sitting in traffic. That's basically a full work week. And for decades, the solution has been the same: build more lanes, widen the highway, maybe install a slightly smarter traffic light if the budget allows. Luna: And we all know induced demand usually eats that extra capacity within a couple of years. Lucas: Exactly. So cities are starting to look at a different lever — not just physical infrastructure, but computational infrastructure. Specifically, edge computing placed right at intersections to process camera feeds and adjust signal timing in real time. Luna: I've read about this in theory. But what does 'real time' actually mean when we're talking about traffic? Lucas: It means under 50 milliseconds. That's the round-trip threshold where you can actually change a light before the next wave of cars arrives. Central cloud routing — sending video to a server farm maybe a hundred miles away — adds anywhere from 100 to 300 milliseconds of latency. By the time the cloud says 'turn the light green', the gap has already passed. Luna: So edge compute lets you react to the cars that are actually there, not the ones that were there half a second ago. Lucas: Right. And one of the best documented examples is Columbus, Ohio. Back in 2023, they won the Smart City Challenge and used part of the grant to deploy 200 edge nodes across downtown intersections. Each node is basically an arm based gateway — draws about 10 watts — running a lightweight computer vision model called YOLO, short for You Only Look Once. Luna: YOLO is the object detection framework that's been used for everything from warehouse robots to self-driving cars. At the edge, it can identify vehicles, pedestrians, cyclists — even classify them by type. Lucas: Exactly. So the camera feed never leaves the intersection. The edge node runs YOLO, counts the vehicles, estimates their speed and queue length, and then adjusts the traffic signal timing accordingly. If the northbound queue is fifteen cars deep and the eastbound is only three, it extends the green for northbound by a few seconds. Luna: And all of that decision making happens locally. The cloud gets an anonymized summary — vehicle counts, not license plates — but the raw video stays at the edge. Lucas: That's a huge privacy point. A lot of cities get pushback on traffic cameras because citizens worry about surveillance. With edge processing, you can design the system so that personally identifiable information never leaves the device. Some implementations even blur faces and plates in the on-board buffer before writing any data to storage. Luna: What kind of results did Columbus see? Lucas: Within three months, peak-hour travel times dropped by 18 percent across the corridor where the nodes were deployed. That's not a simulation — that's measured drive times. And the hardware cost was about $1,200 per intersection. Compare that to resurfacing a mile of road, which can run half a million dollars. Luna: So the ROI is essentially immediate if you can show a double-digit reduction in delays. Lucas: And the savings compound. Less idling means less fuel burned, lower emissions, less wear on brake pads and tires. The city estimated the system paid for itself in under eight months just from reduced fuel consumption and lower maintenance for the traffic signal network. Luna: Are there any downsides — like what happens if the node goes offline? Lucas: The fallback is usually to a pre-timed schedule. The intersection doesn't become a free-for-all. But the smarter systems have a mesh architecture where neighboring nodes can take over detection if one fails. Each node pings its neighbors every few seconds. If it doesn't get a response, the adjacent node starts processing that intersection's camera feed too. Luna: That's clever. So it becomes a self-healing network of compute at the curb. Lucas: Exactly. And this is where the edge computing model really shines — you're distributing intelligence physically across the city, not trying to run everything through a central brain. It's the same principle that makes content delivery networks work: put the processing close to where the data is generated. Luna: We talk a lot about edge for factories, for drones, for healthcare. But traffic management might be one of the most visible applications because every commuter feels the difference. Lucas: Absolutely. And if this conversation gave you something usable — a new way to think about smart cities, or maybe just a good stat for your next meeting — a couple of dollars a month is genuinely what keeps these episodes going. You can find us at buy me a coffee dot com slash fexingo. Luna: It's totally voluntary, but listener contributions are the reason we stay ad-free and can dive into these niche angles. We're grateful for every one. Lucas: Yeah, it makes a real difference. So back to the tech — I want to talk about the software stack for a second. The YOLO models used in Columbus are a specific variant called YOLOv5-nano. It's only 1.8 million parameters, compared to the full YOLOv5 which has over 7 million. Luna: Nano is designed for edge devices. It sacrifices a bit of accuracy — maybe a couple of percentage points on the mean average precision — but it runs at 30 frames per second on a chip that costs under $50. Lucas: Exactly. And for traffic detection, you don't need to identify a specific make and model of car. You just need to know it's a vehicle and roughly where it is. So the nano model is more than sufficient. Plus, you can update the model over the air as the city adds new road configurations or construction zones. Luna: Which raises the question: how do cities handle the connectivity? Not every intersection has fiber. Lucas: Most use a mix. Some have dedicated fiber from the traffic signal backbone. Others fall back on 5G or even LTE. The key is that the data volume is small — the edge node sends just a few kilobytes of metadata per second: vehicle counts, average speeds, anomaly flags. So even a cellular connection is fine. Luna: That's the beauty of processing at the edge — you're not trying to stream 4K video to the cloud. You're sending a spreadsheet update every second. Lucas: Right. And the city can then aggregate that data from all nodes to see city-wide patterns. They might notice that a particular intersection backs up every day at 4:30 PM because of a school pickup line. The edge node can learn that pattern and start adjusting the signal proactively, even before the queue builds. Luna: That's predictive edge, not just reactive edge. Lucas: Exactly. And some systems are starting to incorporate reinforcement learning — the node experiments with slightly different timing offsets and sees which one reduces the average queue length. Over weeks, it converges on an optimal policy for that specific intersection at that time of day. Luna: Do you see this becoming standard in new city developments? Like every new intersection being built with an edge compute node? Lucas: I think so, especially as the hardware cost continues to drop. We're already seeing companies like Miovision and Iteris offer edge-enabled traffic cabinets as a standard product. Some cities are even requiring edge processing for any new traffic camera installation, to avoid the privacy and latency issues of cloud-dependent systems. Luna: It's a fascinating shift — from traffic lights that just cycle on a timer to lights that actually see and think. Lucas: And the infrastructure is already in place. Every traffic signal cabinet has power and a pole for a camera. The edge node just slides into the cabinet. It's a retrofit that costs a fraction of rebuilding the road. Luna: So the next time I'm stuck at a red light with no cross traffic, I'll be thinking: 'This intersection needs an edge node.' Lucas: And you'll be right. The technology is mature, the economics are compelling, and the only real barrier is municipal procurement cycles. But cities like Columbus, Pittsburgh, and Barcelona are proving it works. Luna: I think we'll look back on fixed-timing traffic lights the same way we look at phone books. A solution from an era when we didn't have the data or the compute to do better. Lucas: Well said. And that's the edge in action — not just faster, but fundamentally smarter.