Best practice for automotive AI sensor setup
Cette publication existe aussi en Français
AImotive in Hungary has developed guidelines for sensor setup and data collection for automotive AI systems.
Developing automotive AI for ADAS or automated driving starts with high-quality, precise, and well-documented data enables everything from neural network training, validation and simulation testing.
This data collection can be the most expensive part of the entire AI development lifecycle, requiring vehicles, sensors, drivers, custom software, and countless hours of engineering.
“At aiMotive, we’ve seen too many teams burn through budgets only to discover months later that their collected data isn’t usable. The reasons vary: sensor misplacement, bad calibration, poor version control, or missing diagnostic metadata. This creates a domino effect: development slows, debugging becomes harder, and trust in the data crumbles,” said Zsófia Ülkei, sensor calibration team leader at AImotive. The company, now part of Stellantis, develops AI processor IP and simulation tools for automotive OEMs and Tier One suppliers..
Creating a robust sensor setup
There are key considerations for the sensor configuration before collecting a single byte of data.
- Simulate the sensor layout before physically installing anything. Check for blind spots, overlapping field of view (FOVs), and range limitations.
- Avoid placing GT sensors on moving parts such as side mirrors or trunk lid– this can undermine calibration stability.
- Use rigid mounting structures and remain consistent. One well-planned design should serve or multiple data collection campaigns for cost effective deployment. Avoid heavy, metal structures that have increased weight and associated safety risks, as well as possible shielding and interference effects. aiMotive uses custom AI-optimized, 3D-printed, reinforced mounts.
- Plan for tomorrow’s data needs by using high-quality sensors from the start to ensure that the setup can support evolving and future needs.
- Validate sensor integration and consider which parts of the pipeline will be affected by the addition of new sensors, and ensure to allocate sufficient time for proper validation.
Simulation tools can simulate the sensor setup virtually to validate coverage, optimize placement, and avoid surprises once the car hits the road.
Calibration and synchronization
Sensors are only as good as their alignment and timing. If calibration or sync is off, the dataset becomes a liability.
- Define calibration tolerances based on the use-case. Don’t assume every programme and use-case needs the same accuracy – aim for the most stringent planned use case.
- Continuously check calibration health – ideally with an automated tool that runs during or after every drive.
- Version every calibration change to make debugging traceable and to manage accidental drift with precise validity intervals for calibration data.
- Prioritize selecting hardware-sync-capable sensors wherever possible and include diagnostics to verify timing on the fly.
These checks can be automated to flag anomalies before they become costly problems.
Smart recording software configuration
The hardware that records all the sensor data can fit under the luggage compartment floor, preserving the usability of the trunk, but even the most perfect sensor setup won’t help if the recording software fails.
- Record raw data, not post-processed outputs – this ensures future reprocessing is possible.
- Log everything, not just sensor output, but system state, GPS, diagnostic flags, operator notes, and environmental conditions.
The tools need to control sensor parameters, automatically create metadata, and ensure consistent, traceable data streams from a single interface.
Real-time feedback for drivers
Recording software should equipped with a real-time driver dashboard for monitoring data collection status. This avoids issues where a critical sensor failed halfway through a 500 km route.
- Show live system health on a driver dashboard, so drivers immediately know if something’s wrong. Make sure that the dashboard is clearly visible to the driver and provides straightforward visual clues for quick error detection.
- Implement on-the-fly diagnostics with clear thresholds for flags like sensor overheating, occlusion, or sync loss.
Live sensor diagnostics notify the driver immediately if intervention is required during the drive to ensure that no recording is going to waste.
Optimize data upload and management
Even the best quality data loses its value if it’s difficult to access or filter
- Enable processing pipelines to kick off immediately with automatic uploads. Don’t wait days or weeks for data to reach engineers, deploy upload stations right where vehicles finish their data collection trips.
- Use automatic flags for faster processing, collected during the data collection drives, if a Software-under-Test (SuT) is connected. The SuT knows best when it’s confidence is lower then expected, and these moments can be filtered and prioritized.
- Track diagnostic trends and thresholds over time to spot hardware issues before they escalate.
High-quality data is not a given; it’s a result of rigorous design, disciplined execution, and the right tooling.
“At aiMotive, we’ve seen firsthand how aiData helps OEMs and Tier-1s avoid costly mistakes, speed up development, and build AI that works – in the lab, in simulation, and on real roads,” said Ülkei.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
