An edge-AI driver-safety app that spots distraction in real time.
It watches for the things that pull a driver’s attention off the road, like a phone in their hand or their eyes turning away, and warns them the moment it happens. Because it runs on the chip itself, what the industry calls edge AI, there’s no cloud, no lag, and almost no load on the rest of the car’s computer.
Talk to us →From new chip to working app in three weeks.
Cars are starting to watch their drivers. Are they on their phone? Are their eyes drifting off the road? A warning about that only helps if it arrives in time, and sending the camera feed to the cloud for analysis adds a delay that can make it arrive too late. So the AI has to make the call on the spot, on a chip inside the car, which also keeps the video in the vehicle rather than sending it anywhere.
While Axelera AI makes the chip, rinf.tech’s job was to build the app that runs on it, and to do it on a brand-new chip the team had never worked with. A safety app has to watch several things at once: whether the driver is looking at the road, whether they’re using a phone, whether their head or body has turned away. Each one is a separate AI model, and rinf.tech had to get all of them running together, in real time, on Axelera’s Metis chip.
rinf.tech delivered it in about three weeks. The app runs at full video speed, and it leans on the AI chip so completely that the car’s main processor stays almost free.
From new chip to working app. rinf.tech had never used this chip before the project.
Full, smooth video. 30 frames a second on an HD stream — real-time, with no stutter.
Load on the main processor. The AI runs on the chip, so the device’s own processor stays free for everything else.
Working as one system. Spotting, following and reading the driver, combined into a single app.
Learning a new AI chip from scratch, and building on it in weeks
Axelera needed an application built on its Metis AIPU, its new AI accelerator chip, to prove to partners and customers what the platform can do, and it needed it on a short timeline. That is the challenge rinf.tech took on.
A specialized AI chip is not plug-and-play. Unlike a regular processor, it won’t just run an AI model as it comes: each model has to be converted into the chip’s own format, made to fit, and tuned so the chip itself does the heavy lifting instead of quietly handing work back to the regular processor. rinf.tech had never worked with the Metis chip or Axelera’s Voyager SDK, the software toolkit for putting models on it, so all of that had to be learned from scratch.
Then came the harder part. A driver-safety app is not one model but several, running in a chain: find the driver, follow them, read their face and body, and decide what to do. rinf.tech had to get all of them working together, in real time, with the work staying on the chip, inside a few weeks.
Turning what the camera sees into a decision the driver can trust
A camera feed on its own only gives you detections: a box around a face, a box around a phone. The value is in the judgment on top. Is the driver fine, do they need a nudge, or is this serious? The AI models were defined in Voyager SDK YAML descriptors, and a separate Python application layer read their output and applied the driver-monitoring rules, with short timers so a single stray frame doesn’t trigger a false alarm. The result is one of three states: OK, Warning, or Alert.
Rather than build everything at once, the team added the models one stage at a time, and measured each stage on the chip before moving on.
What it watches
From a live camera or a recorded video, the app tracks whether a driver is present, follows them frame to frame, spots a phone near them, reads their face position and where they are looking, and checks their body posture. If the driver sees a warning but has things under control, a simple raised hand tells the system to stand down instead of escalating.
# 01 — Spotting
Model: YOLOv5s (COCO)The system finds the driver and any phone in the picture.
# 02 — Following
Adds: OC-SORT trackerIt locks onto the same driver and follows them frame to frame, instead of starting over each time.
# 03 — Reading
Adds: RetinaFace (face landmarks) + YOLOv8 pose (body pose)It reads face and body position, so it can tell where the driver is looking and recognise a raised-hand “I’m OK”.
Full speed, with almost no load on the host CPU
The app ran at a full 30 frames a second, and the host CPU stayed under 1.2% the whole time. Those two numbers matter most together: real-time speed alongside an almost-idle main processor is what proves the work genuinely ran on the AI chip, not on the host.
What this proves goes beyond running one model. rinf.tech took a brand-new chip and turned it into a running application in weeks: setting up the hardware and toolkit, getting the models onto the chip, combining them into one pipeline, adding the decision logic and the on-screen alerts, benchmarking each step, and documenting it all.
Held real-time speed at every stage
In practice: even with all four AI models running together, the app held 30 frames a second, the same speed it ran at with a single model. It did not slow down as models were added.
The chip carried the load, not the processor
In practice: as models were added, the main processor’s share of the work rose across the three stages from 0.7% to 0.8% to 1.1%, under half a percentage point in total. Nearly everything ran on Axelera’s Metis chip, not the host processor, so the rest of the system stays free. Power use was not measured.
What the app does
In real time, the app detects the driver, spots a phone, and reads their attention and posture, all on the chip. In short:
From an unfamiliar chip to a running product, in weeks
The project showed that rinf.tech can take a brand-new, unfamiliar AI platform and turn it into a running, benchmarked product in a matter of weeks. The clearest sign the work landed: since the engagement, Axelera has recommended rinf.tech to several of its own clients as an integration partner.
- Industry
- Automotive · in-car driver monitoring
- What it is
- Edge AI
- Built on
- Axelera Metis AIPU + Voyager SDK
- Time to demo
- ~3 weeks (4–6 week window)
- Speed
- 30 FPS on HD video
- Status
- Proof of capability