AI Traffic Cameras: What the Milesight ITS Report Misses

What the report actually says

Milesight has published a 33 page white paper on AI-powered smart traffic cameras, covering market sizing, technology generations, application scenarios, industry challenges and its own product positioning. It is a vendor document and reads like one in places, but the structural argument is worth engaging with.

The central framing is a three generation progression. First generation cameras captured video for a human to review later. Second generation added automatic recognition, principally licence plate recognition (LPR) and optical character recognition, which turned footage into structured records for tolling, access control and enforcement. Third generation cameras run inference on the device: multi-object recognition, event detection and real-time analytics. Milesight summarises this as a move from seeing, to recognising, to understanding.

On sizing, the report leans on a Strategic Market Research forecast of $2.4bn in 2024 rising to $7.8bn by 2030, a 21.3% compound annual growth rate. It splits the market into fixed cameras (over 60% share in 2024, and the dominant category), mobile cameras (smallest share, fastest growth in percentage terms) and integrated ITS cameras (the segment with the highest forecast CAGR). Regionally it puts North America at roughly 35% of deployments, with Asia-Pacific the fastest growing.

Read the numbers with care

Two things stand out on a close read. First, the report cites several different LPR market sizes from different analysts within a few pages of each other: roughly $298m in 2024 from one source, and $1.43bn in 2025 from another. Those are not reconcilable as the same market. They are almost certainly different segment definitions, and the report does not say which. Anyone building a business case from either figure needs to go back to the primary source.

Second, the road fatality figure appears as 1.2 million in one section and 1.19 million in another. Trivial in isolation, but it is a reminder of what these documents are. Market forecasts published by camera vendors are directional, not audited. Use them to understand where a manufacturer thinks demand is heading, not as an input to a capital plan.

Where the report is on firmer ground is the challenges section, and it is unusually candid for vendor material. It concedes that rain, snow, fog and low light degrade recognition accuracy. It acknowledges that plate format variation across jurisdictions defeats single-model recognition. It flags data privacy exposure directly, citing ACLU analysis of automated licence plate reader networks and GDPR requirements on data minimisation, purpose limitation and storage limitation. And it names system integration complexity, incompatible protocols and inconsistent data formats, as a persistent barrier.

The connectivity gap

Here is what the report does not do. It devotes 33 pages to cameras and roughly two paragraphs to the network those cameras depend on. That is a fair reflection of how the industry markets ITS, and a poor reflection of where deployments actually fail.

A fixed enforcement camera at a signalised junction is a permanent, unattended, weather-exposed endpoint that must produce legally admissible evidence on demand. Consider what that implies for the connectivity design.

Bandwidth is a design decision, not a spec sheet line

A camera streaming continuous full-resolution video over cellular will consume tens of gigabytes per month per device and will still not deliver reliable real-time analytics. The entire point of edge AI is that it inverts the traffic profile. Inference runs on the camera, and the network carries plate strings, vehicle attributes, timestamps, event flags and a handful of evidential JPEGs. That takes a well-provisioned deployment from a video backhaul problem to a metadata problem, which cellular handles comfortably.

Design the link for the exception, not the average. Evidence retrieval, firmware updates and post-incident video pulls are the peaks that matter. This is where LTE Cat 6 or 5G headroom earns its cost, and where 5G RedCap starts to look like the sensible middle ground for camera-class endpoints over the next refresh cycle.

Uptime and addressing

An enforcement camera that drops off the network during a violation has not simply lost a data point. It has created an evidential gap. Dual SIM failover across two UK operators is the baseline, not a premium feature, and the failover behaviour needs testing under real cell congestion rather than by pulling a SIM in the office. Where a single site cannot reach a second usable network, a multi-network roaming SIM gives one profile access to whichever operator has usable coverage at that pole.

Addressing is the other recurring failure. Most camera management platforms and video management systems expect to reach the device. On a standard consumer-grade cellular APN behind carrier-grade NAT, they cannot. The clean answer is a private APN with fixed addressing, or an outbound VPN tunnel from the router to a central concentrator. Our guide to IoT SIM types, private APNs and fixed IP covers the trade-offs in full.

Data protection is a network architecture problem

Milesight is right to raise GDPR. In the UK, ANPR data is personal data, and a deployment is subject to the UK GDPR and the Data Protection Act 2018, alongside the surveillance camera guidance that applies to relevant authorities. Data minimisation and purpose limitation are not achieved by policy documents. They are achieved by architecture: filtering at the edge so non-matching plates are never transmitted, encrypting in transit, and terminating the tunnel inside a controlled environment rather than exposing a management interface to the public internet.

That principle, the network as an enforcement point for the data policy rather than a passive pipe, is the same one we set out in the IoT Security Guide.

Where Milesight's routers fit

Milesight is unusual among camera manufacturers in having a credible industrial router line of its own, which makes the omission in the white paper more surprising. The UR series covers most of the roadside cases described in the report, and one model is built specifically for camera backhaul.

The practical pattern for a UK junction deployment looks like this. A UR75 or UR35 in the controller cabinet, PoE PSE powering the camera over a short Ethernet run, dual SIM across two operators, a private APN with a fixed IP terminating an IPsec tunnel back to the VMS, and Node-RED or Python on the router handling any local buffering during a network drop. For a pole with no cabinet, a UF51 moves the router outdoors and reverses the PoE direction.

Milesight also ships DeviceHub for fleet configuration and MilesightVPN for remote access, which matters more than it sounds when a scheme runs to hundreds of intersections. Full specifications are on the Milesight router product pages, and our Routers & Gateways database lists comparable models from other manufacturers.

What this means for UK deployments

The report's forward-looking sections cover V2X integration, autonomous driving co-operation and real-time traffic digital twins. Those are real research directions, and for a UK local authority procuring cameras in 2026 they are largely irrelevant to the decision at hand. What is relevant is that every one of those futures assumes a camera that can push structured data to a platform continuously, securely and cheaply.

That assumption is a connectivity requirement, and it is the requirement most likely to be discovered late. If you are scoping ITS work, treat the camera selection and the backhaul design as a single exercise. Read the Milesight report for the technology direction, discount the market figures accordingly, and spend the saved effort on the network. See our IoT Connectivity Guide for the underlying cellular decisions.