A single camera streams video around the clock, and almost none of it is interesting. A loading bay is empty most of the night; a production line looks the same for thousands of frames between defects. Sending every frame to a central data centre for analysis spends bandwidth, compute and privacy on frames that contain nothing.
The alternative is a tiered design: decide what matters as close to the camera as possible, and escalate only the frames that need a bigger model. This article sets out that design and the trade-offs at each tier.
Small detectors are now good enough to filter
Real-time detection models have improved sharply. RF-DETR, presented at ICLR 2026, reports 48.0 AP on COCO for its nano variant, 5.3 AP above D-FINE nano at similar latency, and its largest variant is reported above 60 AP on COCO while still running in real time. The authors also use architecture search to find accuracy-latency trade-offs for a target dataset without retraining for each configuration.
For edge architecture, this means the first tier no longer has to be a crude motion detector. A small, fine-tuned detector at the gateway can classify most frames correctly and decide which few need more.
A three-tier vision architecture
| Tier | Hardware | Job | Output |
|---|---|---|---|
| Camera or sensor | Camera SoC, smart sensor | Motion, simple triggers, compression | Encoded stream, events |
| Gateway or site server | Mini PC, embedded module, small GPU | Detection, tracking, counting, redaction | Metadata, cropped regions, alerts |
| Regional or central GPU | Data-centre accelerators | Open-vocabulary segmentation, vision-language reasoning, re-identification where lawful | Answers, labels, reports |
Escalation rules decide cost. Typical triggers: a detection with low confidence, a rare class, a rule violation (a person in a restricted zone), or a periodic audit sample to measure the gateway's accuracy.
Send the smallest artefact that answers the question. Metadata (classes, boxes, counts, tracks) is tiny compared with video. A cropped region is smaller than a frame. A frame is smaller than a clip. ONVIF Profile M standardises the communication of analytics metadata and events between edge devices and clients such as video management systems and cloud services, which makes metadata-first designs interoperable.
The takeaway: the gateway's job is to turn video into metadata and send pixels upstream only by exception.
Privacy is an architecture decision
Most camera footage contains people. In the EU, the EDPB guidelines on video devices set out how GDPR applies to video surveillance, including the legal basis, transparency and retention. The AI Act goes further for some uses: Article 5 prohibits the use of real-time remote biometric identification systems in publicly accessible spaces for law enforcement except in narrowly defined cases, and prohibits building facial recognition databases through untargeted scraping of facial images from the internet or CCTV footage.
A tiered design makes compliance easier:
- Redact at the gateway. Blur faces and plates before any frame leaves the site, unless the purpose lawfully requires them.
- Minimise escalation. Fewer frames leaving the site means less personal data in transit and at rest elsewhere.
- Pin the upstream tier. When frames must go to a larger model, send them to a provider in the right jurisdiction, under a policy that forbids retention and training on inputs.
- Retain metadata, not video. Counts and events usually satisfy the business need without keeping footage.
The UK applies equivalent rules under the UK GDPR. Outside Europe, rules differ widely, but the same architecture keeps options open.
Operating a fleet of gateways
Gateways are many, remote and physically exposed. Each should:
- hold its identity in hardware, so a cloned SD card does not produce a second trusted gateway;
- verify every model file by hash before loading, so a tampered detector cannot be swapped in;
- report its accuracy through audit samples, so drift is caught before it costs money;
- update through signed releases.
What this means for edge AI teams and hardware makers
- Budget compute per camera at the gateway, and size escalation capacity for the exception rate, not the frame rate.
- Fine-tune small detectors on your own scenes; general benchmarks are a starting point.
- Design redaction and retention into the pipeline before deployment.
- For hardware makers: a secure element and a signed update path matter as much as TOPS for fleet gateways.
Where Tenzro fits
Tenzro is an open network where machines of every size sell AI work. For vision, built today:
- Vision models on the network. Providers serve detection (RF-DETR n through 2xl and D-FINE), point- and box-prompted segmentation (SAM 2, EdgeSAM, MobileSAM), text-promptable segmentation (SAM 3 and 3.1), image embeddings (CLIP, SigLIP2, DINOv2, DINOv3) and frame-pooled video embeddings. These are the escalation tier, priced per call by each provider.
- Placement and policy. Providers declare a coarse region and jurisdiction. Operators publish signed policies listing the models they serve, their jurisdictions and their data terms (retention days, whether they train on inputs), and serving outside a declared jurisdiction is a provable violation.
- Verified model files. Before weights load, each file is hashed and checked against the catalog's pinned hash and the model's verified manifest; a mismatch refuses the load.
- Gateway identity. Machines are anchored in attestable hardware roots; a device without a TPM can hold its identity key in another machine's TPM.
What is not built: the edge build for small devices forwards model calls to providers rather than running detectors itself. A gateway or mini PC with enough memory can run the standard build, which serves models that fit a CPU. A small-GPU edge variant for Jetson-class modules and a harness-host role for local filtering and escalation are planned. The network is live in preview. See tenzro.com.
