In November, highway concessionaire PLUS Malaysia announced it had rolled out automated number plate recognition (ANPR) to track vehicles passing through North-South Highway toll plazas. The system is meant to ensure motorists are paying their toll fees as they should and is said to be the first step towards the goal of multi-lane free flow (MLFF), a form of open road tolling that dispenses with the usual booths.

Now, technology giant Nvidia has detailed the computer vision and artificial intelligence (AI) tech found in PLUS’ ANPR system, which incorporates the company’s proprietary graphics processing units (GPUs) and software. VehicleTrack, developed by Malaysian startup Tapway, can read a number plate and detect a car’s class, make and colour in just 50 milliseconds – even when it’s travelling at up to 40 km/h.

Founded by former aerospace engineer Lim Chee How, Tapway entered into the project by responding to a call for help in video analytics back in 2019. The client, PLUS, was looking to be able to effectively track a car’s entry and exit points, given that the North-South Highway runs a closed system for tolls.

Specifically, the company wanted to prevent users from using one payment method to enter the highway and another to exit, presumably in preparation of its rollout of Touch ‘n Go’s radio frequency identification (RFID). This was meant to stop motorists from either trying to cheat the system or being double charged. “We showed them how with computer vision — just a camera and AI — you could solve all that,” said Lim.

Tapway trained and ran its AI models using Nvidia A100 and V100 Tensor Core GPUs; the result is a system that works in all light and weather conditions with a consistent 97% accuracy, Nvidia said. Each GPU can manage up to 50 video streams at once thanks to an Nvidia Triton Interference Server, enabling the ANPR system to process up to 28,800 images a minute on edge servers using Nvidia A10, A30 and T4 GPUs.

“Triton is a real lifesaver for us,” said Lim. “We had some scaling problems doing inference and multithreading on our own and couldn’t scale beyond 12 video streams in a server, but with Triton we easily handle 20 and we’ve tested it on up to 50 simultaneous streams,” he said.

According to Lim, Tapway uses a dataset of up to 100,000 images to prepare a new AI model for a customer in just a few hours, rather than several days in the case of a CPU-based system. The company builds its apps using the Nvidia DeepStream software development kit (SDK) and optimises its AI models via TensorRT, an SDK for high-performance deep learning inference.

As we’ve previously reported, PLUS uses the ANPR system not only on RFID lanes, but also Touch ‘n Go and SmartTAG ones. This enables it to track all users, even if they enter using a TnG card or SmartTAG and exit using RFID, charging their eWallet (via RFID) rather than their card. The company has installed 577 cameras so far and plans to grow this number to nearly 900 at 92 toll plazas, Nvidia said.

Even with all this technology, PLUS’ ANPR system is not foolproof. To that end, the concessionaire has implemented a dedicated validation centre where backend staff are able to augment the system by identifying damaged, dirty or non-standard (i.e. with “fancy” fonts and spacing) plates.

Since its founding in 2014, Tapway has implemented 3,000 sensors in 500 locations throughout Malaysia and Singapore. Its AI tech not only detects number plates but also help study consumer buying habits in malls; the company even wants to aid regional carmakers and palm oil producers in improving quality control inspections, Nvidia said.