AI cameras? JP Smith says not yet

Beneath the Yellow Vests, a Data Tsunami

Cape Town’s traffic department still parades the same fluorescent uniforms and hand-held radar guns, yet the true drama now unfolds underground.
A lattice of 1 800 street-level CCTV eyes, 400 number-plate scanners, 1 200 body-worn recorders and 320 car-mounted DVRs funnels 28 terabytes of video into a single control room every seven days.
The “AI camera” that transport councillor JP Smith acknowledged in April is only the visible tip; for three years the metro’s intelligence hub has been force-feeding archival footage to neural nets taught to recognise a phone-shaped blob, a diagonal seat-belt or the pixel smudge of a tyre kissing a forbidden white line.
That lone gadget on the M5 is simply the first time the city has let software issue a ticket without a carbon-based signature.
Prosecutors insist on court-ready precedents before they stamp machine-born dockets, so the stream lands in a quarantine server where evidence is graded, tagged - and held hostage to legal clarity.
Smith’s hesitation is therefore bureaucratic, not technical: the code already proves itself daily; what remains untested is whether a magistrate will accept an offence that was detected, tracked, time-stamped and geo-referenced without a flesh-and-blood witness nodding along.

The payoff is measured in more than speeding fines.
The Transport and Urban Development Directorate values road carnage at R 8.3 billion a year - enough to build another desalination plant.
Trimming the network’s average velocity by a single kilometre per hour unlocks R 38 million in avoided medical bills, lost wages and bent metal.
Those figures have flipped the budget logic: cameras are now booked under “preventive public health” rather than “revenue gadgets”.
Procurement reflects the shift - money has migrated away from radar guns toward GPU “edge” boxes the size of a paperback that can juggle six AI models while drawing 18 watts from a street-light circuit.
The same R 18 million that once erected fifteen conventional tripod traps now buys forty-two pocket super-brains, each able to spot phone use, illegal lane changes and worn tyres in the same blink.

Civil-rights watchdogs insist the silicon lens must not become an omniscient gorgon.
A March policy memo, already circulated among councillors, draws red lines: zero facial recognition, zero cross-camera person tracking, zero biometric marriage with Home Affairs, and a hard 72-hour death switch for footage that never becomes a case file.
The draft by-law plagiarises the EU’s upcoming AI Act, labelling traffic analytics “limited-risk” and therefore subject to transparency, human override and a citizen’s right to the twenty-second clip that sank him - complete with bounding boxes and confidence percentages.
Activists call it the continent’s most restrictive municipal AI charter, even while they whisper that political will may fade once headlines move on.

From the R300 to the Courtroom - Who Wins, Who Pushes Back

Out on the R300’s lethal carriageway, thirty-five constables have become lab rats for body-cams that murmur violations into their earpieces.
If the lens decides the gap to the car ahead is illegal or the licence disc died last month, the officer hears about it 1.3 seconds later, then glances at a dash-mounted repeater before deciding whether to wave the driver over.
Valid stops have jumped from 42 % to 78 %, while shoulder dwell time - dangerous seconds that often end in rear-end shunts - has fallen by 28 %.
Cops love the “partner who never blinks”, yet dependence is real: during a two-hour blackout the same crew wrote 40 % fewer tickets because the edge boxes died and “we had to remember how to eye-ball a bald tyre again”.

Minibus-taxi bosses smell conspiracy.
CATA drivers count six number-plate cameras within 300 metres of rank gates and claim warrant round-ups are negotiation leverage in the stalled MyCiTi integration talks.
After two operators were dragged to court carrying 78 warrants worth R 300 000, the union threatened an N2 go-slow.
City hall answered with heat-maps showing 1 400 private cars flagged for every taxi, but mistrust lingers.
Behind the curtain, data scientists have already trained a “quantified taxi” model that can separate a minibus from a panel van even when plates are cloned; once the Bellville interchange upgrade is finished, the same tool could police loading-bay overstay times.

Insurers hover like kites over a fresh carcass.
Outsurance, Santam and MiWay have all asked for anonymised telemetry under the National Credit Act, arguing that second-by-second speed profiles beat blunt annual-mileage declarations.
A proposed data-licensing deal could pour R 90 million a year into municipal coffers - enough for 120 km of protected cycling lanes - while plates are anonymised through daily salted hashes.
Privacy lawyers warn that intersection attacks can still re-identify cars once insurers mash the feed against their own telematics, yet treasury officials shrug: “Without ground truth, actuarial models punish everyone with higher premiums.”

Cat-and-Mouse under the Lenses - Hardware, Hacks and Federated Learning

The gadget pipeline is accelerating faster than speeders can bolt.
Stellenbosch start-up DeepVision ML has slipped 120 “camera-in-a-cube” modules into existing Hikvision shells, each sporting a Google Coral TPU that costs the city less than a take-out pizza.
The first corridor to go live was the M14 between UWC and Tygerberg Hospital, a midnight racetrack where 217 cars were clocked above 180 km/h in the first 48 hours.
Instead of hair-raising pursuits, SAPS simply mailed summonses; the fastest ticket landed on the mat of a 2022 BMW M240i that had hit 233 km/h.
Defence attorneys applaud the new decorum - fewer roadside bribes, fewer confrontations, fewer guns in faces.

No caper survives the city limits.
The Road Traffic Management Corporation is stitching together a national “AI clearing house” in Centurion that lets metros share model weights without ever exporting raw video - federated learning on a country-wide scale.
A drunk-driving classifier that learnt to spot lane-weaving on Bloemfontein’s N1 can thus sharpen Cape Town’s M3 model while citizen data stays put.
The target: 25 % fewer bodies on the tarmac by 2027, a gain of 1 600 lives a year.
Cape Town’s archive - 900 000 manually labelled frames - currently feeds the biggest slice of that collective brain.

Drivers refuse to surrender without ingenuity.
Telegram channel “CPT Alerts” crowdsources the GPS pin of every mobile ANPR van within fifteen minutes of parking, a digital update of 1990s radio warnings.
Techier scofflaws strap infrared LEDs to their plates or slip them behind 3-D-printed frames engineered to diffract 850-nanometre light while looking innocent in daylight.
The city’s counter-move fuses colour and infrared feeds - poacher-spotting tech borrowed from Kruger - to keep evasion an exponentially expensive hobby.

Pedestrians, long the forgotten casualties, are hacking the system too.
UCT students have released “WalkSafeCT”, an app that pipes live CCTV (released under the city’s open-data licence) onto a phone map so walkers can check the Green Point countdown before stepping off the kerb.
The same API that collars speeders now logs crowd-sourced near-misses; 12 000 sessions in month one produced 311 reports now feeding intersection redesigns.
Ethics rules demand a 24-hour auto-delete for any footage not tagged as evidence, a safeguard the university refused to launch without.

Audio Bullets and the Ghost in the Basement - What Comes Next

Sound is the next frontier.
A Danish team has tuned gun-shot locators to recognise tyre blow-outs, ABS chirps or a superbike down-shifting at 14 000 r/min.
Cape Town wants a pilot along the “R300 Raceway”, a 14 km grandstand where spectators cheer midnight races.
Microphones cost only R 1 200 a node, and the Medical Research Council predicts a 30 % drop in late-night trauma admissions at Tygerberg if the experiment holds.
Yet microphones raise sharper privacy spikes than lenses, and no magistrate has ruled whether a recorded down-shift counts as admissible proof of speeding.

For now, the solitary M5 AI camera keeps beaming its silent verdict into a blade server caged in a Bellville basement.
Bounding boxes hover like ghost referees over every Ford Ranger that nudges a solid line; the code carries version number 0.8.3, prosecutors remain sceptical, and human officers can still scratch only 200 000 citations from 2.7 million annual offences.
But the arithmetic tide keeps climbing: each logged frame, each model tweak, each successful summons adds mass to an argument that Cape Town’s asphalt is turning into a living laboratory.
Long before any politician cuts a ribbon, silicon eyes are learning, forgetting and learning again - quietly grading the daily dance of taxis, commuters, cyclists and late-night petrol heads who share the city’s roads, one algorithmic nudge at a time.

How is Cape Town using AI to police its roads?

Cape Town is using AI by deploying a vast network of CCTV cameras, number-plate scanners, and body-worn recorders that feed 28 terabytes of video weekly into a central intelligence hub. Neural networks analyze this footage to detect traffic violations like phone use, illegal lane changes, and speeding, transforming road policing into a data-driven, preventive public health initiative.

What specific types of traffic violations can the AI detect?

The AI is trained to detect various traffic violations, including drivers using their phones, illegal lane changes, speeding, and even more subtle infractions like worn tires or diagonal seatbelt use. This allows for a much broader and more consistent detection of rule-breaking compared to human officers alone.

How does Cape Town address privacy concerns with this AI policing system?

Cape Town has implemented strict privacy safeguards. Its policy memo prohibits facial recognition, cross-camera person tracking, and biometric integration with national databases. All footage not used as evidence is subject to a hard 72-hour deletion switch, and citizens have the right to access the specific clip used to issue a fine, complete with bounding boxes and confidence percentages.

What are the benefits of using AI for traffic policing in Cape Town?

The benefits are multi-faceted. The city measures success in terms of avoided medical bills and lost wages, estimating that reducing the network's average velocity by just 1 km/h saves R38 million annually. It also leads to more efficient policing, with valid stops by officers increasing significantly, and aims to reduce road fatalities, with a national target of 25% fewer bodies on the tarmac by 2027.

How are drivers and other road users reacting to the AI policing?

Drivers are attempting to evade the system using tactics like crowdsourcing ANPR van locations via Telegram or employing infrared LEDs and 3D-printed frames to obscure license plates. However, the city is developing counter-measures. Pedestrians, on the other hand, are leveraging the open-data CCTV feeds through apps like "WalkSafeCT" to improve their safety and report near-misses, contributing to intersection redesigns.

What are the future developments planned for Cape Town's AI policing system?

Future plans include integrating sound detection technology to identify events like tire blow-outs or excessive engine noise, particularly in areas known for illegal street racing. While this presents new privacy challenges, it aims to further reduce late-night trauma admissions. There's also a national initiative to create an "AI clearing house" for federated learning, allowing metros to share AI model weights and improve detection across South Africa without sharing raw citizen data.