How Edge Computing Powers Real-Time Expressway Tolling

Lucas: Luna, I don't know about you, but the last time I sat in a toll booth lineup, I sat there for about four minutes watching a guy count change by hand. And I thought — this is a solved problem. Why are we still doing this? Luna: Right? It feels like something out of the nineties. But actually, moving tolling to the edge is a surprisingly complex engineering challenge. Lucas: Exactly. And that's our angle today. We're going to look at how edge computing makes all-electronic, barrier-free tolling possible — specifically on the Mumbai-Nagpur Expressway in India. It opened in 2022, six hundred and one kilometers, and it's one of the first major expressways designed from day one for full edge-based tolling. Luna: Okay, so first — what's the old model? I mean, we've had toll tags for years. ez pass in the US, electronic toll collection. Why isn't that already 'edge'? Lucas: Good question. Traditional electronic toll collection still relies on a centralized backend. When your car passes a reader, that tag ID gets sent to a central server — maybe in a data center hundreds of kilometers away — which verifies your account and deducts the fee. That round trip can take anywhere from a hundred milliseconds to a full second, especially during peak hours when the server is hammered. Luna: And a second doesn't sound like a lot, but when you've got thousands of cars per hour, that lag means cars have to slow down, queues form, and you still need boom gates to prevent people from driving through without payment. Lucas: Precisely. The Mumbai-Nagpur Expressway took a different approach. They deployed local compute nodes — essentially mini data centers — right at each toll gantry. These nodes run the vehicle detection, tag reading, license plate recognition, and account lookup locally. The transaction completes in under thirty milliseconds. That's fast enough that cars can maintain highway speed. No gates, no slowing down, no bottleneck. Luna: So the edge node is basically a hardened server sitting in a roadside cabinet, running a stripped-down database of active accounts? Lucas: That's exactly right. They're running what's essentially a real-time in-memory cache of the tolling database — maybe the last ninety days of transactions and a whitelist of active tags. The node syncs with the central system periodically, but for each transaction, it doesn't need to phone home. And if a car doesn't have a tag, the same node runs an automatic number plate recognition model — an AI model — to read the license plate and cross-reference it against a vehicle registration database. Luna: That's a lot of processing in a box with no air conditioning, sitting on the side of a highway in the Indian summer. How do they handle the heat and dust? Lucas: Yeah, environmental hardening is a big part of it. The cabinets are passively cooled, sealed against dust and water ingress, and they use solid-state storage — no spinning disks. The compute modules are off-the-shelf industrial PCs running Linux, but they're ruggedized. And power is a concern too — each cabinet has a battery backup and can run for several hours on its own if the grid goes down. Luna: And this matters because… if a node goes offline, you can't just stop the highway. You'd have a revenue leak. Lucas: Exactly. The system is designed with redundancy — each gantry has two nodes in active-passive failover. If one fails, the other takes over within milliseconds. The total system availability target is 99.99 percent. And they're achieving it, according to the operator Mahindra Highway. Luna: So what does this mean for the economics of the road? I've heard that traditional toll collection eats up something like 15 to 20 percent of toll revenue just in operating costs — staff, maintenance, cash handling. Lucas: Right. The Mumbai-Nagpur Expressway claims a 40 percent reduction in operating expenses compared to a conventional toll plaza. No cash handlers, no boom gate maintenance, fewer staff. Plus, because cars don't stop, fuel consumption drops, emissions drop, and throughput increases. The expressway handles about fifty thousand vehicles per day, and during peak hours, it processes over fifty thousand transactions per hour per direction. Luna: Fifty thousand per hour. That's about fourteen transactions per second. And each one needs to be accurate — you can't charge the wrong car. Lucas: Right. Accuracy is critical. The ANPR model has to be robust to different lighting, weather, dirt on plates. They trained it on millions of Indian license plates — which have a specific font and color scheme. The model runs on an NVIDIA Jetson module at the edge. They report a plate recognition accuracy of over 98 percent, and for RFID tags it's essentially 100 percent. Luna: So what happens if the model misreads a plate? Or a car with no tag and an obscured plate goes through? Lucas: Good catch. The system doesn't just let it go. It captures an image of the vehicle and stores it locally. That image gets sent to a central enforcement system, and a human reviews it later, cross-referencing with traffic camera databases. If it's a violation, a ticket is mailed to the registered owner. The edge node itself doesn't make enforcement decisions — it just flags exceptions. Luna: So the edge handles the real-time transaction, but the non-real-time enforcement still happens in the cloud. That's a smart split. Lucas: Exactly. It's the classic edge-cloud architecture. Low-latency, high-reliability tasks at the edge; batch processing and analytics in the cloud. And that brings up another interesting point — the edge node also generates a lot of data. Every transaction, every image, every performance metric. That data is compressed and sent to the cloud periodically for analytics — traffic patterns, usage forecasting, maintenance alerts. Luna: So the cloud side can detect, say, that gantry seventeen's node is running a couple degrees hotter than normal, and schedule a maintenance visit before it fails. Lucas: Precisely. Predictive maintenance is a huge value add. They're already using that data to optimize power usage — adjusting the compute load based on traffic volume. At 3 a.m. when traffic is light, the node can throttle down, saving energy and extending hardware life. Luna: I want to ask about security. These nodes are sitting on the side of the road. What prevents someone from tampering with them or injecting fake transactions? Lucas: That's a serious concern. The cabinets are tamper-proof — if someone opens the door, a sensor triggers a remote alert and the node wipes its cryptographic keys. All communication between the node and the central system is encrypted, and the node has a hardware security module to store signing keys. Each transaction is digitally signed, so even if someone intercepts the data, they can't forge a transaction. Luna: And the RFID tags themselves — are they encrypted? Lucas: Yes. The tags use a challenge-response authentication protocol. The node sends a random number, the tag encrypts it with a secret key, and the node verifies the response. That prevents cloning. It's the same kind of security used in contactless payment cards. Luna: So the whole system is built like a distributed banking network, but on a highway. Lucas: That's exactly the right analogy. And it's not just India. Similar systems are rolling out in Norway, in Australia, in parts of the US. But the Mumbai-Nagpur Expressway is one of the first greenfield deployments — built from the ground up with edge computing, not retrofitted. Luna: And honestly, if today's episode gave you a new way to think about edge infrastructure — beyond the usual factory floor or data center use case — that's the kind of thing we love to hear. If it was worth a coffee to you, you know the link: buy me a coffee dot com slash fexingo. Lucas: Yeah, it really does help keep the show ad-free and focused on the technical details. And speaking of details — I want to circle back to something that surprised me when I was researching this. Luna: What's that? Lucas: The total number of edge nodes on the expressway. Six hundred and one kilometers, with gantries approximately every ten kilometers. That's about sixty gantries, each with two nodes. So roughly one hundred and twenty edge servers, all running 24/7, processing millions of transactions per day. And the entire system was built for about twelve million dollars — hardware, installation, and software. Compare that to the cost of a single traditional toll plaza, which can run five to ten million dollars just for construction. Luna: So the payback period is surprisingly short. And the ongoing savings are substantial. Lucas: Right. And that's the pattern we're seeing across edge computing — once you shift real-time decision-making to the edge, you unlock cost savings and performance gains that centralized architectures can't match. Tolling is just one example, but it's a really clear one. Luna: So what's next? Do you see this becoming the standard for all new highways? Lucas: I think so. The technology is proven. The cost is coming down. And with connected vehicles and autonomous driving on the horizon, having a real-time digital infrastructure on the roadside becomes even more critical. The edge node that reads your toll tag today could someday be the node that communicates with your car's driving system about road conditions ahead. Luna: That's a pretty compelling vision. And it starts with a box on the side of the road that can read a license plate in thirty milliseconds. Lucas: Exactly. The edge is where the physical and digital worlds meet. And tolling might be one of the most visible examples of that intersection.