AI to Advance Vision Zero
Dane Manary

Measure streets, not just crashes

In September, Mayor Zohran Mamdani released his strategy to produce safer streets, called Vision Zero Reimagined. It opens with a letter from the mayor: “No New Yorker should lose their life while walking, biking or driving in our city…” What follows is a plan that runs 108 pages, including more than 100 policies. At the press conference in Queens for the plan’s release, Mamdani was asked for his views on the decline in enforcement of traffic laws by the NYPD. He replied that “I do not measure success by an increased number of tickets, or by an increased number of violations. To me, the judgment of success is by a change in behavior.”

To make that judgment, the City needs to actually be able to measure the behavior of those who use the streets. Unfortunately, on the plan’s own account, it can’t do that yet. AI can help — if we let it.

The City already knows a great deal about life on the street. Its speed and red-light cameras log violations, and the cameras on its buses log who blocks bus lanes, bike lanes and bus stops. 311 collects residents’ complaints while police reports and hospitals note crashes and injuries. But these records sit in separate systems, and each one counts the violations or harm, but not necessarily the trips and the vehicles that produce them. Page 43 of Vision Zero Reimagined concedes that “per-trip risk data would offer a more precise understanding of trends, particularly given the rapid expansion of motorized two-wheeler use over the past decade.” It commits to solve that problem by “expand[ing] traffic counting capacity to include more areas and all modes.” It doesn’t say how.

The good news is that, with all this data, the City already knows how; the sensors necessary are, in part, already on New York’s streets, though the plan never mentions them. The City should enrich them with AI, so that it uses those sensors to gather, collate and understand data at a scale no team of analysts ever could. That would give the City a record to act on. 

Sensors on the streets

From July 2023 to August 2024, the City’s Office of Technology and Innovation and the Department of Transportation (DOT) ran a $100,000 pilot that installed 12 roadside computer-vision sensors across four boroughs. Anything the cameras saw, they classified: car, bus, truck, bicycle or pedestrian. They recorded speeds, dwell times and, most usefully, conflicts: the negotiations and near-misses that didn’t become incidents, and as such would not otherwise have been recorded. DOT kept the sensors running after the pilot, increasing coverage to about 20 sites; in June, Commissioner Mike Flynn announced the department would extend the program to about 100 across all five boroughs, to “design safer streets and encourage safer behavior.” Those sensors would join a much larger network that DOT already runs, though those hundreds of traffic cameras across the city merely watch traffic rather than count it.

Strikingly, Vision Zero Reimagined seems ignorant of this program. The plan certainly knows about the cameras: speed cameras, red-light cameras and the bus-mounted Automated Camera Enforcement system appear throughout. But they are envisioned only as enforcement tools rather than as measurement tools. The phrases “computer vision,” “artificial intelligence” and “machine learning” do not appear, and the two mentions of sensors describe equipment bolted to trucks, not street furniture.

Of course, since the pilot began in 2023, the world has changed. Until recently, by DOT’s own description, staff collected traffic counts “manually by reviewing video.” One civic-minded resident, Benjamin Arnav, working on his own with stills from DOT’s traffic cameras, trained a model to spot mopeds riding illegally in bike lanes. As he reported here in Vital City, he found at least one every five minutes on the Queensboro Bridge bike path. And that was nearly three years ago. Today’s vision models can watch every frame from a hundred poles at once, and can be taught a new category without new hardware. 

That last one matters: Whereas the 2023 pilot counted every bicycle as a bicycle, today’s can distinguish an e-bike from a pedal bike from a moped. If new transport tools come along, the models can start tracking them by learning from only a few thousand labeled examples. The frontier of knowledge has extended outward. 

So has the frontier of understanding. The hundreds of traffic and enforcement cameras DOT already runs produce more imagery than any planning team (that the City can afford) could ever watch. AI can close that gap: An analyst might ask, in plain English, which blocks saw near-misses rise after a loading zone went in, and get back the answer, the evidence and the dataset, with all the work shown. The same models read text even more easily than video footage, so the City’s other records, like 311 complaints, crash narratives, violations logged by bus cameras and more, can be included in the analysis. 

In other words, the “data warehouse” that Vision Zero Reimagined promises is within reach, sooner and at less cost than we could have foreseen even a few years ago.

It isn’t news that street tensions in New York flare up at the curb, where delivery vans, street couriers on e-bikes and mopeds, pedestrians, and city buses compete for space. Package deliveries rose from 1.1 million a day in 2017 to 2.5 million in 2024, by the City Comptroller’s count, to say nothing of other incidents, like pedestrians being hit by scooters. Vision Zero Reimagined has an answer to this, which is the new Office of Curb Management at DOT, announced in April. Its aim is to take “a holistic approach to organizing curb space” and reduce “double parking and parking in bike and bus lanes.”

The office has yet to name a director or attach a budget line, which might encourage a view that the City doesn’t take it seriously. Let us instead give the benefit of the doubt, and say its mandate is being determined. If so, then part of that mandate should now be clear: It should have oversight over the current and future sensors, and the analysis that flows from them. It should identify the blocks where our existing data already point to being hot spots for violations and roll cameras out there first, and follow with cameras whenever a corner is built out or a loading zone installed. Then, the office can deliver on the mayor’s promise, and report whether conflicts fell, and by how much.

Automated enforcement is already on the books

AI could not just enable far better strategic decision-making about street design, e-bike policies and more. It could enable far more effective enforcement of the rules when people break them and cause chaos.

Since 2023, state law (Vehicle and Traffic Law § 1111-c-1) has let DOT run camera enforcement of four curb rules: blocking a bus stop, bike lane or bus lane, and double parking. In practice, that has meant putting cameras on Metropolitan Transportation Authority buses. But the statute authorizes cameras “that may be stationary or mobile,” at “locations determined by such department of transportation,” and so far as I can tell, none of them enforce the curb. Speed and red-light cameras enforce only their own violations, and the number of red-light cameras is in any case capped by the state at 600, a limit they’ve reached. By DOT’s own account, its fixed bus-lane cameras only catch drivers moving in bus lanes, and curb violations are left to the cameras on buses. The plan lists such enforcement as something to “explore,” and the obvious place to do that exploration is in the sites that the curb office will be monitoring anyway. The mayor doesn’t want to measure success in police stops; cameras offer a way to enforce Vision Zero without them. 

All that would be required is a simple tweak from Albany. As written, the four rules permitted for automated enforcement don’t include abuse of loading zones. Two bills already in committee in Albany, A5440A from Assemblyman Steven Raga and S8665 from state Sen. Julia Salazar, would add these to the mix. The bills deserve the City’s support.

Some may disagree, on the grounds that this proposal amounts to laying the foundation for a surveillance state. That worry is already present in the debate over automated license plate readers, which can log every passing car… which means that, over time, they can map where people go. The answers to that problem, short retention periods and strict limits on who may see the data, apply here too. AI sharpens the problem, since an AI sophisticated enough to identify a model of e-bike is also capable of identifying its rider. In this regard, New York is a few moves ahead. The statute that authorizes camera enforcement, Section 1111-c-1, bans facial recognition and every other form of biometric identification on the images “for any purpose.” The rest is a matter of keeping two jobs apart. Measurement work doesn’t need to identify anyone, and indeed the City’s already-installed counting sensors do not; they classify what they see on the device, discard the video, and send along only the tallies. With the video discarded after processing, nothing is left to identify anyone. A policy tweak would help as well: The City should extend the statute’s limits on enforcement cameras to include measurement sensors also.

Conversely, enforcement work does require that monitoring systems be able to identify bad actors. That’s why New York ties identification to the license plate and forbids the rest. It’s a start, but it does leave a loophole big enough to ride a moped or an e-bike through: Those vehicles often don’t have plates and aren’t registered, meaning violations can’t be automatically enforced.

The chain of action that will make the mayor’s test into action is clear. Install cameras and sensors to measure. Set up AI models to interpret. Report the findings, to know what is working, and sanction recklessness and abuse, through dispassionate automated enforcement. The record of fines and penalties will record what the City is doing, and over time will show whether matters are improving. 

Mayor Mamdani said tickets are not how he measures success. He’s right: A better measure of success is fewer injuries and deaths, and streets with less congestion and fewer near-misses. To get there, New York will have to combine the tools it already has — cameras, sensors, automated enforcement authority — with the newly available tool of AI, and the curb management office, which at time of writing has still not been organized. All of these, taken together, will give the City a new opportunity to make the streets safer for everyone.


Great! You’ve successfully signed up.

Welcome back! You've successfully signed in.

You've successfully subscribed to Vital City.

Success! Check your email for magic link to sign-in.

Success! Your billing info has been updated.

Your billing was not updated.