Edge AI & Sensors

Buying advice

Choose field hardware for its service life

Prototype on whatever is quickest, but choose the field hardware with its whole service life in mind. A device bolted to a pole has to cope with heat, cold, vibration and power interruptions for years, and replacing it means a site visit. Ask each maker for the rated operating temperature, the expected production lifetime of the part and how much notice it gives before end of life, so a design that works at one site can be rolled out to the next. By October 2026, a buyer could spend as little as about $80 on the Ambarella CV5 AI Vision SoC or as much as about $6,000 on the Ouster OS1-128 LiDAR, and all of them showed a price.

Govern the models you build yourself

When your team trains and deploys its own detection models, you take on work a camera maker would otherwise do. NIST's AI Risk Management Framework offers a structure for mapping, measuring and managing risks such as false alarms, missed detections and accuracy that drops in rain, glare or darkness. Keep records of the training data, test results and model versions running at each site. If you package these components into a product sold in the EU, Regulation (EU) 2024/2847, the Cyber Resilience Act, adds cybersecurity duties, including handling vulnerabilities and providing security updates.

LiDAR and thermal sensors need their own checks

Laser scanners such as the Ouster OS1-128 and Cepton Vista-X120 are laser products, so ask for the laser class assigned under IEC 60825-1:2014 and keep the label and safety information with your installation records. Point clouds and thermal video are heavy streams, so size the edge computer for the sensors you may add later, not only the first one. Agree how timestamps from cameras, LiDAR and thermal units will be synchronised, because fused detections fall apart when clocks drift. Plan for cleaning and recalibration too, since dust, spider webs and knocks degrade sensors on exposed poles.

What to look for

• AI performance, power draw and the camera streams each device can handle
• The software stack and model formats your developers already use
• Range, field of view and resolution for LiDAR and thermal sensors

Common questions

Why process video at the edge?

Running analytics next to the camera reduces the video sent across your network and can keep footage on site. It suits remote locations and sites with limited bandwidth.

Is this hardware ready to use out of the box?

Most items here are development and integration components. Expect to pair them with cameras, software and enclosures, or to work with an integrator.

Sources and further reading

• AI Risk Management Framework — National Institute of Standards and Technology (NIST).
  A voluntary framework for mapping, measuring and managing risks in AI systems, with companion playbooks.
• IEC 60825-1:2014 Safety of laser products - Part 1: Equipment classification and requirements — International Electrotechnical Commission (IEC).
  The laser safety standard that assigns hazard classes, the basis for eye-safety labels on LiDAR units.
• Regulation (EU) 2024/2847 (Cyber Resilience Act) — European Union, EUR-Lex.
  EU cybersecurity requirements for hardware and software with digital elements, including update duties.