Roadside inference · traffic operations · evidential systems

Smart roads & edge-AI traffic management systems

Traffic-management systems engineered for mixed road use, intermittent backhaul, constrained roadside power and the evidential duties of public authorities.

A road system is not made intelligent by sending every camera stream to a cloud dashboard. On a regional corridor, cellular backhaul may be expensive, variable or absent exactly when an incident matters. The operational boundary therefore starts at the roadside unit: frames are processed locally, only authorised events, signed evidence packages and aggregates are transmitted, and the site continues a defined local mode while disconnected.

The perception problem is also local. A useful taxonomy must distinguish the road users and objects that actually share East African carriageways: passenger cars, heavy goods vehicles, buses, matatus or minibus taxis, boda bodas and other motorcycles, bicycles, pedestrians, handcarts and livestock where the site requires them. Model acceptance is reported per class and operating condition. A single accuracy number can conceal failure on the smallest, most vulnerable or most frequently occluded classes.

The service joins perception to traffic operation. Depending on the mandate and approved scope, that can include adaptive signals and corridor coordination, queue and incident detection, wrong-way or stopped-vehicle alerts, public-transport priority, weigh-in-motion, overload workflow, toll reconciliation, travel-time analytics and road-asset condition monitoring. It does not transfer enforcement power to a supplier: the responsible authority owns policy, review, penalty, appeal and legal-metrology decisions.

Operational design targets

Parameters are set for the site—not presented as past performance

The values below are configurable design ranges or authority-set inputs. Acceptance values require representative traffic, device, power, link and evidence assumptions.

7–12
Detection classes supported
Illustrative taxonomy design range; final classes follow the site decision and local dataset.
50–250 ms
Edge inference latency band
Illustrative design budget on the selected device; not an achieved benchmark.
24–168 h
Offline buffer duration
Configurable sizing range derived from event rate, evidence size, storage, power and outage case.
Authority-set
Evidence retention window
A separate approved schedule applies to transient media, analytics and case evidence.