October is trade fair season for baby monitors. The Hong Kong Electronics Fair autumn edition ran from 13 to 16 October, and the first phase of the Canton Fair from 15 to 19 October. Many booths showed monitor platforms with a local screen and a phone app on the same system, and AI features ranging from cry detection to sleep tracking. A few advertised breathing or heart-rate detection without contact. For a brand walking the aisles, the difficult part is working out how much of what a booth shows is ready to put under your own name.
A booth demo shows that a feature can work once, in conditions the exhibitor chose. To judge a platform, separate three questions. What kind of detection is it: an event such as crying, an inference such as a sleep score, or a physiological measurement such as breathing rate. What stage is it at: booth demo, take-away sample, pilot batch, or shipping product, each of which calls for different evidence. Who owns what: the model, the SDK, the app and any cloud service the features depend on. Vital-sign features need separate handling. In a September 2025 safety communication, the US FDA said infant monitors that claim to measure heart rate, oxygen saturation, breathing rate or temperature are not general wellness products and need FDA marketing authorization. For a brand, adding such a feature is a regulatory decision before it is a product feature.
§01Three kinds of “AI” on a baby monitor booth
Booth signage tends to put every smart feature under the same “AI” heading. For evaluation, they fall into three groups that need very different evidence.
| Type | Examples | What has to be shown | Where the line is |
|---|---|---|---|
| Event detection | Crying, sound level, motion, a person entering the frame | How often it misses events, and how often it alerts when nothing happened, under stated conditions | Usually a consumer feature. Wording matters when an event is framed as a safety outcome. |
| Behaviour inference | Sleep and wake periods, sleep summaries or scores | What the inference is based on (motion over time, sound, both), and how it was checked against a reference | A summary of observed activity is different from a health assessment. Keep the wording on the observation side. |
| Physiological measurement | Breathing rate, heart rate, oxygen saturation, body temperature | Accuracy against a reference method, across infants and conditions, plus the regulatory status in each market | In the US, the FDA treats infant monitors claiming these measurements as devices needing authorization. See §05. |
TABLE 01 · Some event features, such as alerts for a covered face, sit close to safety claims. How they are worded matters as much as how they work. See our article on health-claim language.
The groups matter because a supplier’s demo of all three looks the same at a booth: a screen, a sleeping doll or an adult volunteer, and an alert that appears on cue. The evidence behind each one is very different.
§02Four levels of evidence
Before asking whether a feature is accurate, ask what stage the platform is actually at. Each level proves something different, and a platform can be at a different level for each of its features.
Booth demo
Shows: the feature can run, in conditions the exhibitor chose.
Ask: Is the detection running on the camera in front of you, on a phone, or on a server? Is the screen live or a recording? Can you clap, cover the lens or walk through the frame to see how it reacts?
A sample you can take away
Shows: the feature runs on hardware you control.
Ask: Can it run with no internet connection? Then test it yourself: your own rooms, at night under infrared, with a real crib, blankets, a second child or a pet in the house.
Pilot batch
Shows: the feature survives production variation.
Ask: A test report giving missed-event and false-alert rates, the data it was tested on, and the conditions where it fails. A firmware and model version number tied to that report.
Shipping under any brand
Shows: the platform has been through certification and real use.
Ask: The FCC ID or other certification records of a product that ships on this platform. Update history since launch. How field complaints about the AI features are handled.
FIG. 01 · Many platforms are at L3 or L4 for their basic video features and at L1 for their newest AI features. Ask about each feature separately.
Our article on false alerts and alert fatigue explains why. When the event being watched for is rare, even a small false-alert rate can mean that most alerts are false.
§03Questions about the AI itself
On the camera, on the phone, or on a server? If it runs on a server, the feature works only while the monitor is online. See on-device AI vs cloud.
On a platform with both a local screen and an app, do the AI alerts appear on the local screen when there is no internet, or only in the phone app?
Was it tested on infrared black-and-white video, at the camera height and angle your product will use?
Missed events and false alerts, each stated with the data set and conditions used. One headline accuracy figure on its own says little.
Who retrains or updates the model, how updates reach devices, and whether a brand can hold back an update it has not tested.
Can individual features be switched off for markets where you do not want to offer them, without separate firmware builds?
§04Who owns the pieces
An AI feature on a hybrid platform usually depends on more than the camera. It may need the supplier’s app, its cloud service, a licence for the model, or all three. Before a feature goes into your specification, find out what you would be tied to.
- Model licence: is the AI feature priced into the hardware, charged per device, or charged per year?
- Cloud dependency: does the feature stop working if the supplier’s cloud service changes or ends?
- App and accounts: who controls the app listing and user accounts. Our article on white-label app ownership covers this in detail.
- Video and training data: is any video sent off the device, and if so, is it used to train models? This has to match what your privacy policy tells parents.
- Firmware: who maintains it after launch. See firmware update support.
If the answer to any of these is unclear, the feature is not ready for your brand, however well it demonstrates.
§05Vital-sign features are a regulatory decision
What the FDA said in September 2025
On 16 September 2025, the FDA issued a safety communication advising against unauthorized infant devices that claim to monitor vital signs. It names heart or pulse rate, oxygen saturation, breathing or respiratory rate, and body temperature. It states that these monitors are not covered by the FDA’s general wellness policy and need FDA marketing authorization, and it tells consumers how to check for an authorized device in the FDA’s databases.
The communication is addressed to the US market. Other markets have their own rules for when a product that measures vital signs counts as a medical device.
For a brand, this changes what a vital-sign feature on an ODM spec sheet means. Before adding it, find out:
- Whether any marketing authorization exists for the platform, in whose name, for which device and for which intended use. Check the FDA’s public databases yourself rather than relying on a supplier’s statement.
- What your own brand would need to do to sell that feature in each target market.
- Whether the platform lets you ship without the feature, and how the marketing will be kept free of claims the product cannot support.
Our earlier articles cover the background: how to read cleared, approved, registered and compliant, where a baby monitor ends and a medical device begins, and regulatory status by country.
§06From booth to RFQ
AI platform evaluation · after the fair
Fill this in for each platform you are considering, one feature at a time.
- Classify each feature as event detection, behaviour inference or physiological measurement.
- Record the evidence level for each feature, from L1 to L4, based on what you were actually shown.
- Request a take-away sample and test it in your own conditions, including at night.
- Ask for missed-event and false-alert rates with the data set and conditions behind them.
- Confirm where each feature runs and whether it works with no internet connection.
- List every dependency: model licence, app, cloud service, account system.
- For any vital-sign feature, check regulatory status in your target markets before it goes into the specification.
- Write the marketing claim for each feature now, and check that the evidence supports it.
How we approach it. When a brand evaluates one of our camera platforms, we would rather they test it on their own recordings than rely on our demo. We give clear answers on which parts of a platform are ours and which come from third parties. Our Camera & Video Monitoring page covers the platform and app projects we work on, and the development process shows where sample testing sits.
This article describes how to evaluate any platform. It does not assess any specific exhibitor or product.
Reference: U.S. Food and Drug Administration, “Do Not Use Unauthorized Infant Devices for Monitoring Vital Signs: FDA Safety Communication,” 16 September 2025. Read the FDA communication.
§07Questions buyers ask
How can I tell if an AI baby monitor feature shown at a trade show is ready for production?
Ask what level of evidence exists for that feature: a booth demo, a sample you can take away and test, a pilot batch with a test report, or a product already shipping under some brand with certification records. Many platforms are production-ready for basic video and at demo stage for their newest AI features, so ask about each feature separately.
What test data should an ODM provide for AI cry or motion detection?
Missed-event and false-alert rates, each with the data set and conditions used to measure them, including night-time infrared video. A firmware and model version number should be tied to the report so you know what was tested.
Can a private-label baby monitor include contactless breathing or heart-rate detection?
That is a regulatory question before it is a product question. In a September 2025 safety communication, the US FDA said infant monitors claiming to measure heart rate, oxygen saturation, breathing rate or temperature are not general wellness products and need FDA marketing authorization. Check what authorization exists, in whose name and for which device, and what your brand would need in each market.
Do AI features on a hybrid baby monitor work without the internet?
It depends on where the detection runs. If it runs on the camera, it can work on a local screen with no internet. If it runs on a server, it needs a connection. Ask the supplier, then test it with the internet disconnected.
What should a brand own or control when using an ODM’s AI platform?
Find out who controls the model licence, the app and user accounts, any cloud service the features depend on, the firmware update process, and whether video leaves the device and how it is used. If any of these is unclear, the feature is not ready to ship under your brand.
Comparing platforms after the fair?
Send us the feature list you are considering. We will tell you which features we can support on our platforms, at what stage, and which ones need a regulatory decision first.
Send Your Feature List → Camera platforms · SDK and app projects · Sample testing · OEM / ODM