Attending ISC West? Get Your Free Ticket and Meet Us There!
Get a Free Ticket to ISC West
At FF Group, we believe real-world observations are as valuable as lab results when it comes to advancing recognition technology. Earlier this year, our Chief Product Officer, Jan Hazlbauer, spent two weeks in Sri Lanka. What began as a personal trip quickly turned into a field study: Jan could not resist looking closely at the license plates, vehicles, and traffic mix on Sri Lankan roads.
What he found is a compact case study in why regional adaptation matters for both license plate recognition and vehicle make/model recognition. His reflections highlight the nuances of Sri Lanka’s traffic and offer insights into how recognition systems can be better trained for real-world conditions.
Below, Jan shares his field notes from Sri Lanka:
I spent two weeks in Sri Lanka and, because of my job, I couldn’t stop looking at license plates and vehicle brands on Sri Lankan roads. What looks simple at first—clean plate templates and familiar character sets—turns out to be nuanced for both license plate recognition (LPR) and vehicle make/model recognition (MMR).
When you look at the Sri Lankan license plate templates at first glance, they might look easy to detect and recognize, but there are nuances that make the situation more complicated. For MMR it is also complicated from the first glance, although it depends how deep you plan to go with vehicle analysis. So, what is specific, complicated, and maybe also easy on Sri Lankan LPR and MMR?
The first thing you notice is a typical Asian traffic mix: a huge amount of motorbikes and TukTuks. A TukTuk is a small car which looks like a motorized rickshaw, a three-wheeler covered by a tent cloth. They are very colourful and customized, but in general always the same. Our MMR is not recognizing TukTuk as a class or make for now. Based on my observation, the fleet here skews heavily toward two and three-wheelers - TukTuks and motorcycles dominate, while fewer passenger cars appear on the road, at a share roughly comparable to trucks and buses. Official statistics align with this: as of Q3 2023 Sri Lanka had about 4.85 million motorcycles, 1.18 million three-wheelers, and 0.90 million motor cars registered nationwide (Department of Motor Traffic), and a Western Province study reports ~51% motorcycles and ~20.2% three-wheelers in that regional fleet.
This mix makes the situation easier in one sense, because most vehicles can be classified as “motorcycle” or “TukTuk.” What makes it harder is that the brands and models represented in Sri Lanka are different from Europe or the US. There are many Indian brands like Tata or Mahindra, and a local (in partnership with India) bus and truck brand Lanka Ashok Leyland (LAL). The rest of the represented cars are mostly Asian—Japanese Toyota, Mazda, Honda, Suzuki, Nissan and Mitsubishi—and for trucks/lorries you see Isuzu and Hino. European cars appear—Audi, BMW, Mercedes—and sometimes Land Rover or Peugeot, but they are very rare. From the emerging EV vehicles, BYD is present via an official distributor (John Keells CG Auto), Perodua has re-entered through local partners, and Tata EVs are handled locally by DIMO (Tata’s authorized partner in Sri Lanka).
In general, adding TukTuks and some local Asian and Indian brands could increase the recognition rate significantly, but there is one more aspect: many cars are customized and many are very old and sometimes damaged. This makes MMR accuracy lower in practice. Even within TukTuks you would see different makes and models. Leaders in Sri Lanka are Piaggio with the Ape model (which means “bee” in Italian—yes, the same company that makes the Vespa) and Bajaj with the RE line; others such as TVS and Atul are also present. Whether it makes sense to differentiate TukTuk models for MMR is still a question.
The current system has been active since 2000 (format AA1234), with an update in 2013 (format AAA1234). On Sri Lankan roads, you commonly encounter white front plates and yellow rear plates, which matters for detection and segmentation; multiple Sri Lankan research efforts use the yellow rear-plate prior to locate plates. The small province code (for example WP for Western Province) appears under the national emblem on the left of the plate, and it is not part of the main plate text.
These plates are the most used and seen on Sri Lankan roads, but there are many special and old ones. There are governmental plates (police and others) with Sri Lankan Sinhala letters, inverse plates with digits only, and older plates with embossed letters (including styles with unusual or italic fonts). Old vehicles are not forced to change plates to new ones, so the variety stays high. With that comes another problem: old plates are very often damaged, scratched, and not so reflective anymore—issues that reduce read rates in real traffic. (Local ANPR studies also note practical challenges such as occlusion, lighting, and motion blur, which match what you see on the road here.)
One more issue is solvable on the algorithm level but still might be an issue in practice. The tiny letters (most common is WP) indicating the province of registration are not part of the plate text but are quite visible, and they can confuse the algorithm if you do not mask that zone before OCR. The traffic is sometimes so dense and chaotic that it is very often hard to capture all the plates, especially in free-flow scenarios. I also noticed most of the Suzuki models used here have the license plate not centered in the front, but closer to the right side—yet another small quirk to handle. Good thing is motorcycles have front plates, but of course they are small and often appear in complicated angles, which makes front reads on bikes harder than rears. (Sri Lankan projects frequently emphasize rear-view detection for this reason.)
Sri Lanka illustrates a broader truth: license plate and vehicle recognition are never universal, they are regional sciences. What looks like a simple dataset of clean plates and common brands quickly unravels into a living ecosystem of aging fleets, local manufacturers, and hybrid numbering systems. The lesson is not only to add TukTuks or Indian brands, but to build recognition models that absorb the context of a market—plate formats, vehicle age, traffic density, even cultural choices in customization. When models are tuned to those realities, accuracy stops being a theoretical benchmark and becomes a practical tool for law enforcement, traffic management, and smart-city operations.
The path forward is systematic: study each road environment as carefully as a researcher studies a lab sample, then translate those findings into the model. That is how LPR and MMR evolve from generic software into infrastructure-ready technology.
.svg)