cleoanka

an interactive page · 5G and AI

Opening a lane when it matters

When a camera sees an incident it asks the network for priority. Try what that does to latency in a crowded 5G cell.

canvas & code, no dependencies · RoadGuard on GitHub


1

Opening a lane in a crowded cell

A road camera sends its video to the control centre over 5G. When the cell is quiet, everything is fast. But in evening traffic hundreds of phones share the same base station, packets wait in the queue, and latency climbs to hundreds of milliseconds. Video arriving late at exactly the moment of a crash defeats the whole point of the system.

CAMARA is an initiative where operators expose network capabilities through standard APIs. With the Quality on Demand (QoD) API an application can request a better service profile for one flow, temporarily. RoadGuard uses it like this: the moment the model detects an incident it calls QoD and the camera flow moves to the priority queue; when the incident ends the priority is released.

FIG. 1 — A packet-level cell simulation. Gold packets are the camera, grey packets everyone else. Raise the background load, then press “incident detected”: the QoD call takes effect after a short signalling delay. Right, end-to-end latency of camera frames; the dashed line is the 150 ms target.

Below capacity the two modes are indistinguishable. As load nears capacity the best-effort queue swells and latency explodes; in the priority queue camera packets jump the crowd, so latency stays low. The price is shown honestly: everyone else’s queue grows a little longer at the same time. That is why priority is requested only for the duration of an incident, not all the time.

2

Between false alarms and misses

Driver-behaviour detection (phone use, no seatbelt, looking around) produces a confidence score every frame, and that score jitters. If crossing the threshold in a single frame were enough for an alarm, the system would beep at every hand movement. Raising the threshold cuts false alarms but starts missing real events. The way to improve both is time: a rule like “in at least k of the last n frames” filters out brief flickers and catches behaviour that actually lasts.

FIG. 2 — Green bands are the true “phone use” intervals, the line is the model’s per-frame confidence. Tune the threshold and the temporal filter; below are the intervals where the system raises an alarm and the counts: caught, missed, false alarms.

With the filter off, finding a good threshold is nearly impossible. Turn on a rule like 5 of 8 frames and both false alarms and misses fall across a wide range of thresholds. The price is a few frames of delay, usually worth paying for a safety system. RoadGuard reports its behaviour classes with temporal rules like this and per-class tuned thresholds.

qod
The CAMARA API that requests a better service profile for one flow, temporarily.
queueing
As load nears capacity, latency grows not linearly but explosively.
k / n
Temporal filter: removes flicker, keeps lasting behaviour; costs a few frames.