Science1 publisher2 min readPublished
MIT's city lab estimates New York emissions from 331 traffic camera feeds
A new book from MIT's Senseable City Lab argues that computer vision can quantify streets at a scale pen-and-paper urbanism never reached, and the account of it gives privacy and fairness one line.
The Scientist · Science desk
What happened
- MIT's Senseable City Lab used machine learning to identify vehicle types appearing in 331 New York City traffic cameras and estimated the emissions coming from each automobile in the frames.
- The method appears in a new book, "How AI Sees the City: Urban Visual Intelligence," written by three Senseable City Lab researchers and a remote sensing professor at Peking University.
- Routledge published the book this month, placing computer vision in a line of visual urban study that runs back through photography to Roman marble maps.
- The authors list snarled traffic, dangerous features of intersections and which parts of plazas and parks attract the most people among the questions the technique can be pointed at.
- The MIT account says that across cities more images bring more data and more insight along with more concerns about privacy and fairness.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- capability A city can now get per-intersection emission estimates from camera feeds it already owns and pays for, including at intersections that never got a fixed air monitor.
- constraint Accuracy is bounded by the classifier and the emission factors behind each vehicle class, so buying more cameras buys coverage and leaves the per-vehicle error where it was.
- exposure Everyone crossing a monitored intersection becomes a subject in a traffic and emissions dataset, and the rules on retention and consent fall to whichever agency operates the feed.
- decision A planner choosing between satellite canopy data and street-level imagery is choosing between two different quantities, and a programme that raises one can leave the other unchanged.
A traffic camera measures light. Getting grams of pollutant out of one takes two inferential steps: a classifier assigns each vehicle in the frame to a type, and an emission factor for that type turns the count into an estimate [1]. Each step carries its own error, and the second compounds the first. A van misread as a car draws the wrong factor, and a van read correctly is still scored with its class average, whatever its engine age or load. The MIT News account of the study does not report a check of the camera estimates against measured pollutant concentrations [16].
The scale argument depends on coverage. "Everybody has been observing the urban environment and trying to get some insight," Martina Mazzarello, a research scientist at the lab, said. "But what if we can do that at a large scale and get some insight everywhere?" [8] The estimate describes the intersections those 331 cameras watch. Extending it to the rest of New York is extrapolation, and where the cameras sit sets how far it can stretch.
Fabio Duarte, the lab's associate director, put the method in one line: "With computer vision techniques, each image is a dataset." [3] He also said, "We have to be careful about it." [4] In the MIT account, that caution runs a sentence, and the privacy and fairness concerns there are attributed to rising image volume [7].
The greenery chapter is the cleanest case of instrument choice. Satellite imagery gives tree cover and green space area. Street-level images from phones and other sources measure a different quantity, how much greenery a person glimpses in everyday life, and the book connects that to reported wellness [12]. A plan can lift one of those numbers without touching the other. "The real promise of visual AI is not simply that computers can look at millions of images," said Fan Zhang, the book's co-author at Peking University's Institute of Remote Sensing and GIS [9].
The lineage claim is modest and I think it holds. Kevin Lynch, a former MIT professor, published "The Image of the City" in 1960 and shaped how a generation of scholars read urban form [11]; the Senseable City Lab has been working with urban data since 2004 [10]. "Kevin Lynch at MIT was only using paper and pen," Duarte said [13].
What the 331-camera study demonstrates is throughput: a pipeline that can classify fleets at many locations at once and price their emissions from a table. Demonstrating accuracy would take a paired comparison against reference monitors at the same intersections.
What to watch
- Whether the Senseable City Lab publishes a paired comparison of its camera-derived emissions against reference monitors at the same New York intersections.
- Whether any transport agency adopts the method in procurement, and what retention and access rules it writes for the camera feeds it already runs.
- Whether the street-level greenery measure is tested against health or wellness outcomes at the same locations, or only reported as a correlate.