What makes an edge AI camera hard to manufacture
An edge AI camera combines an image sensor and lens, a system-on-chip with an image signal processor (ISP) and a neural processing unit (NPU), memory and storage, networking and power, in a housing that has to manage heat and often weather. Each of those parts exists off the shelf. The difficulty is that they interact. Model accuracy depends on image quality, which depends on lens focus and per-unit calibration. Sustained inference produces heat that can throttle the SoC or shift focus. And every unit needs a secure identity and update path provisioned on the line.
Most programs start from an ODM camera platform or a chip vendor’s reference design. Whether that platform fits depends on three questions: can it run your model at the frame rate you need, under sustained load, at the hottest ambient you specify; can the factory hold focus and image quality from unit to unit; and can it provision and update devices securely enough for your markets and customers. A demo on a development kit answers none of them.
The decisions that lock the design
| Decision | What to settle before tooling |
|---|---|
| SoC and NPU | Does your model run at the target frame rate after quantization, with every operator supported on the NPU? What is the sustained power at that load? How long will the chip be supplied, and does its origin meet your customers’ procurement rules? |
| Image sensor | Resolution against pixel size and low-light needs; wide dynamic range for backlit scenes; global or rolling shutter for fast motion; near-infrared sensitivity for night mode |
| Lens | Field of view, aperture and distortion; glass or plastic elements, since plastic lenses shift focus more with temperature; infrared correction for day-and-night use |
| Memory and storage | Room for the model, video buffers and two firmware images for safe updates; memory prices and the availability of older DRAM generations have been volatile |
| Network and power | Wi-Fi, Ethernet or cellular; the Power over Ethernet class against the combined load of SoC, IR illumination and any heater (802.3af delivers about 13 W at the camera, 802.3at about 25 W) |
| Housing | Indoor plastic or outdoor die-cast aluminum; ingress and impact ratings, sun load and condensation management |
NPU performance figures quoted in TOPS are not comparable across vendors. What matters is operator support in the vendor’s toolchain, accuracy after INT8 quantization and measured frames per second on your model. Layers the NPU cannot run fall back to the CPU and can cut throughput sharply. Run your own model on the candidate platform before you compare quotes.
Core manufacturing processes
- PCBA assembly with a ball grid array (BGA) SoC and memory, inspected by X-ray. Outdoor products often add conformal coating.
- Camera module assembly in a clean environment, since dust on the sensor or filter is one of the most common optical defects. The lens is focused and fixed with adhesive, either by threading and gluing or by active alignment, in which a machine positions the lens in several axes against a live image before curing.
- Focus verification with a spatial frequency response (SFR) chart after the adhesive cures, because curing can move the focus.
- Enclosure assembly: gaskets, a breathable vent membrane to manage condensation, thermal pads from the SoC to the housing and, for dome cameras, foam isolation between the IR LEDs and the lens so infrared light does not reflect inside the dome and wash out the image.
- Leak test for sealed housings, usually by air pressure decay.
- Provisioning: firmware, secure-boot keys and a unique device certificate for each unit, with every record reconciled.
- Burn-in under load, to catch early failures in memory and power circuits.
- Image quality station: white balance, color, lens shading, defective pixels, IR-cut filter switching and IR illumination, measured against charts under controlled light.
Testing and validation from EVT to PVT
EVT proves the system: model accuracy and frame rate on the target SoC with the real sensor and lens, thermal behavior under sustained inference at maximum ambient, night performance with IR, end-to-end latency, and power against the PoE budget.
DVT repeats the thermal work in tooled housings and adds environmental and compliance testing: ingress protection to IEC 60529 (for example IP66 or IP67), impact resistance to IEC 62262 for vandal-resistant domes (IK10 is the usual claim), temperature cycling, humidity and condensation, UV and salt-fog exposure for outdoor housings, surge and ESD on cable entries, and focus stability across the temperature range. Re-check model accuracy on DVT units, because ISP tuning and lens changes since EVT shift what the model sees.
PVT confirms the line: active-alignment yield, focus capability (Cpk) against the SFR limits, image station correlation with a golden unit, burn-in failure rate, provisioning throughput, and an over-the-air update on production units. To test inference on every unit without customer data, point the camera at a fixed chart or scene and compare its output with a known result. The EVT, DVT and PVT guide covers the stage gates in general.
Security, privacy and certification
| Market or buyer | Typical requirements |
|---|---|
| United States | FCC equipment authorization; product safety certification, commonly to UL 62368-1, which retailers and enterprise buyers expect; California’s connected-device security law (SB-327), under which a unique password per device or a forced change at first use counts as a reasonable security feature. Government and many enterprise buyers require NDAA Section 889 compliance, and the FCC restricts new equipment authorizations for companies on its Covered List. |
| European Union | CE marking under the Radio Equipment Directive for wireless models, including its cybersecurity requirements (assessed against the EN 18031 standards) since August 2025, or under the EMC Directive and the applicable safety rules for wired models; RoHS, REACH and WEEE. The Cyber Resilience Act’s vulnerability-reporting duties began in September 2026, and its main requirements apply from December 2027. |
| United Kingdom | UKCA or CE, plus the PSTI Act: no universal default passwords, a published route for reporting vulnerabilities and a stated security-update period. |
| Japan, Korea, Australia | Radio and EMC approvals (MIC, KC, RCM). |
| Interoperability | ONVIF profile conformance if the camera must work with third-party video management systems. |
Cameras that identify people create privacy obligations for whoever deploys them, and in the EU the AI Act restricts some biometric uses. Get specialist advice on the intended use before marketing claims are written.
Main cost drivers
- SoC and memory, usually the largest share of the bill of materials.
- Sensor and lens, especially low-light sensors and glass lenses.
- Housing: die-cast aluminum, optical-grade domes, coatings and gaskets for outdoor models.
- IR illumination, heaters for cold climates and PoE circuitry.
- Active-alignment machine time, focus yield and burn-in hours.
- Certification and security testing, repeated for regional variants.
- Software: video codec licensing (H.265 has more than one patent pool), update infrastructure and security maintenance for the support period you commit to.
What to ask a prospective OEM/ODM
- Which SoC platforms have you taken into production, and can we benchmark our model on one now?
- Who does ISP tuning for our sensor and lens, and how many tuning rounds are included?
- Do you use active alignment, and what is your focus acceptance limit and measured capability?
- What clean-room class or clean-bench controls apply to camera module assembly?
- How are secure-boot keys and device certificates generated, injected and protected?
- How long is burn-in, and what failure rate do you see?
- Can you document the origin of the SoC, sensor and camera module for NDAA or similar buyer requirements?
- How long will you support security updates on this platform?
Common mistakes
- Benchmarking on a development kit with a fan and no housing.
- Training the model on images from a different sensor and lens than the product uses.
- Leaving IR reflection inside the dome to be fixed after tooling.
- Choosing a chipset whose origin rules out the customers you plan to sell to.
- Promising a security-update period without deciding who will maintain it.
- Treating condensation as an assembly problem instead of a design problem.
Where Zefinity helps
We help AI hardware teams test whether an ODM camera platform or reference design can carry their model, prepare the product for manufacturing, and find OEM/ODM partners with proven camera module, thermal and provisioning capability. If your product is a module for someone else’s machine, see the computer vision module guide. Learn more about our work in AI hardware.
Frequently asked questions
Why does my model perform worse on the production camera than on the development kit?
The production camera usually has a different sensor, lens, image signal processor tuning or quantized model, and each changes what the network sees. Collect training and validation images through production-intent optics and the final image pipeline, quantize and test on the target NPU, and re-check accuracy on DVT units, because lens and tuning changes after EVT are common.
What does NDAA-compliant mean for a camera?
It usually refers to Section 889 of the US National Defense Authorization Act for fiscal 2019, which bars US federal agencies from buying video surveillance and telecommunications equipment from certain named companies and their affiliates, or equipment that uses it as a substantial or essential component. Buyers ask for it to confirm that the chipset, camera module and firmware do not come from those sources. Check the SoC and module origin, not only the brand.
Do edge AI cameras need a fan?
Many do not. Most fixed cameras are cooled passively through the housing, which suits sealed outdoor enclosures. Whether that works depends on the SoC's sustained power at your model's load, the highest ambient temperature and sun exposure you specify, and the housing material. Measure sustained frame rate and SoC temperature in a tooled housing, not on a development kit with a fan.