In January 2025, New York City began charging vehicles to enter Manhattan’s Congestion Relief Zone (CRZ), the first program of its kind in the United States. This project measures the policy’s effect directly from the street. A computer vision pipeline counts vehicles in footage from the city’s public traffic cameras, and traffic in the same November 14 to January 4 window is compared before and after the policy, with anomalous periods such as holidays excluded from the baselines.
Method
Each camera contributes instantaneous vehicle counts from object detection. The counts are aggregated into hourly averages across a typical week, so rush-hour peaks, weekday and weekend patterns, and the before-and-after difference can be compared at every camera and mapped across the city.
Results
- Traffic fell inside the zone. Peak observed car count per frame dropped 15.8% at cameras within the CRZ.
- Traffic also fell outside the zone. Cameras outside the CRZ recorded a smaller 10.9% drop in peak observed car count per frame.
- Changes are resolved camera by camera. The interactive map reports each camera’s change as a percentage or an absolute count, for the whole week, weekdays only, or weekends only.
Live view
The analysis is ongoing, with regular updates to track how traffic evolves under the policy over the long term. A companion real-time map shows the camera network as the pipeline processes it.
Traffic integrity and edge-cloud security
The camera pipeline behind this study is also the sensing layer for our work on the integrity and security of city-scale traffic systems.
- Verifying crowdsourced traffic data. Navigation platforms infer congestion from phone GPS traces, which an attacker can fake with emulated “ghost” vehicles. Digital Eyes on the Road (USENIX VehicleSec 2026) cross-checks GPS-based traffic levels against this pipeline’s per-camera vehicle counts, produced every 15 seconds from NYC DOT cameras. In Manhattan it detects 91.5% of moderate injected attacks and about 99% of stronger ones, and where a location has no camera of its own, neighboring cameras alone retain a consensus congestion signal at 96.68% of the locations tested.
- Worst-case attacks on edge-cloud systems. Camera networks like this one are natural edge-cloud workloads: edge sites process frames near the cameras and offload to the cloud when load spikes. Worst-Case Attacks in Reactive Edge-Cloud Systems (IEEE INFOCOM 2026) proves that expected latency and the probability of SLA violations stay convex in the size of an attack that takes edge servers offline, even with reactive offloading, so the most damaging attacks concentrate on a few critical sites. In simulation, these attacks cause up to 40% more damage with the same resources.
Limitations
Camera-based vehicle counts are a proxy for traffic, not a direct measure of travel times or congestion. The current pipeline includes stationary vehicles, which can raise measured density on streets with heavy parking, and it measures aggregate flow without separating individual lanes or travel directions.