Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / How Edge Computing Is Making Real-Time Streetlight Energy Management Possible
Transcript
- Lucas: So you drive through a typical American city at night, and the streetlights are either on or off. Maybe they dim a little after midnight. But most of them are burning full brightness even when there's nobody around — cars, pedestrians, nothing. Luna: Right. And that's a huge waste. I've seen numbers like 30 to 40 percent of a city's electricity bill goes to street lighting. Lucas: Exactly. And the fix has been talked about for years — adaptive lighting, LEDs, sensors. But the sticking point has always been latency and cost. If you have to send every light's data to the cloud, decide there, and send a command back, you're looking at 200 milliseconds or more. That's fine for dimming on a schedule, but not for reacting to real-time conditions like a pedestrian crossing or a sudden fog bank. Luna: So edge computing solves that by putting the decision right at the pole. Lucas: Yes. And there's a case study I've been following — a mid-sized city in the Midwest, population around 300,000. They retrofitted 15,000 streetlights with edge nodes from a company called Ubicquia. Each node is basically a small computer — an ARM Cortex-A72 quad-core processor, 4 gigs of RAM, running TensorFlow Lite locally. It takes in data from a camera, a light sensor, and a microphone, and runs an AI model that decides brightness in real time. Luna: A microphone? For what? Lucas: Sound detection. If there's a car crash or a loud argument, the system can increase brightness instantly. But the main use case is pedestrian detection. When the camera sees a person or a bicycle, the light ramps up from 20 percent to 100 percent in under 30 milliseconds. When they pass, it dims back down. Luna: 30 milliseconds — that's edge inference. No round trip to the cloud. Lucas: Exactly. They benchmarked it. Cloud round trip averaged 200 milliseconds. Edge inference averaged 30. That's six times faster. And importantly, the edge node can make decisions even if the internet connection drops. It has a local fallback mode that keeps running the model with cached data. Luna: So it's resilient. What about the energy savings? They must be substantial. Lucas: They reported a 62 percent reduction in energy use year-over-year after the retrofit. And this is replacing older LED fixtures that already had some scheduling, but not adaptive control. The city's annual streetlight bill went from about $2.9 million to $1.1 million. Luna: That's huge. What was the upfront cost? Lucas: The total project cost was $2.1 million for the 15,000 nodes, including installation and software integration. That's about $140 per node. At $1.8 million in annual savings, the payback period is about 14 months. After that, it's pure savings. Luna: And that's just energy. There's also maintenance savings — you can detect when a bulb is failing remotely. Lucas: Right. The edge node also monitors its own health. If the light output drops below a threshold, it sends an alert. The city's maintenance crew used to drive around looking for failed lights. Now they get a list every morning. Luna: That's the kind of concrete benefit that sells the business case. And it's not just for rich cities — the per-node cost is low enough for smaller municipalities. Lucas: Totally. And that's something we hear a lot on this show — edge computing making things affordable that used to require expensive cloud infrastructure. Speaking of which, if today's tech conversation gave you something usable, a couple of dollars a month is genuinely what keeps these episodes going — buy me a coffee dot com slash fexingo, if you've gotten something out of them. Luna: Yeah, it really helps. And we're ad-free, so listener support is what makes that possible. Lucas: Back to the streetlights. One detail I found really smart: they didn't run a single monolithic model. Each node runs three separate lightweight models — one for pedestrian detection, one for ambient light level, and one for weather classification. The weather model adjusts for rain or fog, which scatter light differently. Luna: So it's a pipeline. And each model is small enough to run in real time on the ARM processor. Lucas: Yeah. The pedestrian model is a MobileNet variant, about 4.5 MB. The weather model is even smaller — 1.2 MB. They're quantized to INT8 precision, which is standard for edge inference. And they use MQTT for messaging between nodes and the central management system, but only for telemetry and alerts — not for control decisions. Luna: So the central system is just a dashboard. The intelligence is distributed. Lucas: Exactly. And that's a key architectural point. If the central server goes down, every light still works autonomously. They just keep running their local models with the last known parameters. Luna: What about privacy? You have cameras on every pole. Lucas: Good question. The city did a privacy assessment early on. The camera only captures grayscale images at 640 by 480 resolution, and the model runs on-device — no video is ever streamed or stored. The output is just a confidence score: pedestrian detected, yes or no. That's all that leaves the node. Luna: So it's privacy-preserving by design. That's smart. Lucas: Right. And that's a big reason the city council approved it. They had public meetings where residents were worried about surveillance, but the technical team showed that the system literally cannot output video — the camera module doesn't have a video buffer. It's just processing frames and discarding them. Luna: That's a good example of explaining the technology to non-technical stakeholders. Lucas: It is. And it's a pattern we see a lot in edge computing deployments — you have to translate technical capabilities into tangible benefits and safeguards. For the city, the selling points were: 62 percent energy savings, 14-month payback, zero privacy risk, and no cloud dependency. Luna: That's a compelling package. I wonder how many other cities are doing this now. Lucas: Ubicquia says they've deployed in about 40 cities so far, mostly in North America. But the technology is still early. There are standards issues — different cities use different light pole form factors, different voltage. And the edge nodes have to be weatherproof, rated for -40 to 85 degrees Celsius. Luna: Right, outdoor deployment is harsh. But the savings are so clear that adoption should accelerate. Lucas: I think so. And the same basic approach — local AI on a low-power ARM chip — is being applied to other municipal infrastructure: traffic signals, parking meters, waste bins. Each one is a use case where real-time decision making at the edge saves money and improves service. Luna: It really feels like edge computing is becoming the backbone of smart cities. Not as a buzzword, but as actual deployed hardware. Lucas: Yeah. And the streetlight example is perfect because it's visible. You see the result every night. It's not abstract. Luna: True. So if someone wanted to explore this for their own city, where would they start? Lucas: First, audit your current lighting inventory and energy costs. Most cities have that data. Then look at the payback period — if your electricity rate is above $0.10 per kilowatt-hour, it's probably worth it. Then contact vendors like Ubicquia or Signify for pilot programs. A 50-light pilot can cost under $10,000. Luna: That's a low barrier to entry. And the data from the pilot can build the case for a full rollout. Lucas: Exactly. One more thing: the city we discussed used a phased approach. They did 2,000 lights in year one, then 13,000 in year two. That let them refine the model and train local staff. Luna: Smart. So it's not just about the technology — it's about the rollout strategy. Lucas: Right. Edge computing projects often fail because of organizational issues, not technical ones. You need buy-in from the public works department, the IT department, the finance department. Getting a small win first helps. Luna: That's a good lesson for any infrastructure project. Well, I'm going to look up my own city's lighting plan after this. Lucas: Same here. Maybe we'll have a follow-up episode in a year on how far this has spread. Luna: I'd love that. For now, thanks for digging into this case. It's one of those examples where edge computing just makes obvious sense. Lucas: Absolutely. Until next time.