MIT Traffic turns roadside cameras into a traffic survey that never ends: virtual tripwires drawn on the frame count and classify every vehicle, calibrated cameras measure real speeds in km/h, and speeding, wrong-way and lane violations are captured with photo evidence and a human review queue. It also watches the road as it runs — stopped vehicles and queues raised as they happen, corridor journey times, and signal timings checked against your own measured demand. No loops cut into the carriageway, no radar, no survey crews.
Draw a tripwire anywhere on the camera frame and it counts from that moment. Endpoints are stored as fractions of the frame, so changing resolution or swapping the camera never moves your count line.
Every crossing is typed — car, van, truck, bus, motorbike, bicycle — and directional, so you get composition and turning movements, not one anonymous number.
Mark a rectangle of known size on the road and the camera learns the ground plane. Speeds come from a least-squares fit over roughly a second of motion, not the gap between two frames.
Where a camera can't be calibrated, pair two lines with the distance between them and time vehicles over the section. Every pass records which method produced its speed.
The system projects a metre grid back onto your camera image — if the grid lies flat on the road, the calibration is right. Impossible geometry is rejected when you save it, not discovered in the data months later.
Watch vehicles cross, typed and timed, in real time — commissioning a site is a few minutes of looking at the screen, not a week of waiting for data.
Volumes by hour, day, direction and vehicle class, with flow rates, volume/capacity ratio and a Level-of-Service grade per segment.
Speeding, wrong-way, illegal stopping and restricted-lane use are detected automatically; red-light joins them wherever a signal-phase feed is available.
Nothing is issued by a machine alone. Every violation lands in a queue with its evidence frame, its measured speed and its calibration record, to be approved or rejected — one at a time or in bulk.
Over-limit passes store a frame; ordinary passes don't. That's the difference between a few hundred images a day and a few hundred thousand.
A vehicle stopped on the carriageway, a queue building, a car going the wrong way — raised within seconds, held open with a running duration, and cleared automatically when the road does.
Every road graded free-flowing to congested from the speed traffic is actually doing. A road at capacity and a road that has broken down both carry high flow — only one of them is still moving.
Match the same plate at two points and you have what a corridor really takes end to end, its reliability, and an origin–destination matrix showing where traffic goes next.
Enter the timings a junction runs and your measured demand says whether the green is shared out the way traffic arrives — or whether the cycle is simply longer than the demand needs.
Model segments, their cameras, their speed limits and their zones, so counts and violations arrive already attached to a place on your network.
Filter violations and passes and export them for enforcement, consultants or a council report — the raw data is yours, on your servers.
| The traditional way | MIT Traffic | |
|---|---|---|
| Getting counts | ✕Book a crew, lay pneumatic tubes, wait a fortnight, receive a PDF | ✓Every vehicle, every day, already running — the survey never stops |
| Installation | ✕Cut inductive loops into the carriageway; close a lane to do it | ✓Mount a camera, or reuse one you already have, and draw a line on the frame |
| Moving a count point | ✕A new deployment, a new crew, a new invoice | ✓Drag the line to somewhere else on the frame |
| Classification | ✕Inferred from axle spacing, and only roughly | ✓The AI names the vehicle type on every single crossing |
| Speed | ✕A radar box measuring one point on one lane | ✓Any calibrated camera, plus section speed over a known distance |
| Evidence | ✕A number on a device, and an argument about it | ✓An evidence frame, the measured speed and the calibration behind it |
| Who decides | ✕Whatever the box recorded | ✓A person reviews every violation before it goes anywhere |
| How current your data is | ✕A survey from eighteen months ago | ✓A live dashboard, and the full history behind it |
| Coverage | ✕The two or three points you could afford to survey | ✓Every camera on the network, at no extra hardware cost |
| Knowing a lane is blocked | ✕A phone call from whoever is stuck in it | ✓An incident on screen seconds after the vehicle stops moving |
| Journey times | ✕A consultant's floating-car study, a handful of runs | ✓Every plated vehicle that drives the corridor, continuously |
| Signal timings | ✕Set once at commissioning, reviewed when someone complains | ✓Checked against the demand your own cameras measured this hour |
Virtual counting lines, per-class and directional counts, live detection, and flow analytics with level-of-service grading.
Live todayGround-plane calibration with a verifiable metre grid, velocity-fit speeds in km/h, and paired-line section speed where calibration isn't possible.
Live todaySpeeding, wrong-way, illegal stopping and restricted-lane detection, evidence frames, a human review inbox with bulk actions, and export.
Live todayStopped vehicles, queues and wrong-way movements raised the moment they appear and cleared the moment they end, with a per-road threshold so a signal queue never reads as a breakdown.
Live todayCorridor travel times and an origin–destination matrix built by matching the same plate at two points — where traffic actually goes, not where it was surveyed.
Live todayYour own measured flows turned into signal-timing advice: whether the green is shared out the way demand arrives, and whether the cycle is longer than the demand needs.
Live todayCrossing the stop line against a red. The detection is built; it needs a live signal-phase feed from the controller, which is a per-junction integration rather than a camera change.
On the roadmapEvery phase above marked shipped is running in production today, on the same cameras and the same calibration — nothing on this list required new hardware on the road once the first camera was pointed at it. Red-light is the one capability that depends on something outside the camera: a live phase feed from the signal controller.
Bounding boxes jitter by a pixel or two, and near the horizon one pixel can be a metre — so a two-frame speed is noise. MIT Traffic fits a velocity over roughly a second of tracked ground positions. Measured against ground truth: 1.7% error at 30 km/h, 2.3% at 60.
Every other camera-speed product asks you to trust a number. This one draws a metre grid back onto your road: if the squares lie flat, the geometry is right. Impossible quads are refused at save time rather than quietly producing plausible nonsense.
Calibrated and section speeds have different error characteristics, so every pass records which one produced it. Nobody has to guess whether a number is defensible.
Flow is computed from the raw crossings on demand, so correcting a line's position re-derives history instead of leaving a stale summary that disagrees with the detail.
Automatic detection, manual decision. That's the design that survives a challenge — and the review queue is built in, not a spreadsheet someone maintains on the side.
A stopped-vehicle alarm that fires on every car at a red light is worse than none — operators stop looking. The threshold is set per road and has to clear a whole signal cycle, and a vehicle must have been moving before stopping counts, so parked cars never fill the log.
An incident is a condition, not an alarm: it opens, carries a running duration, and closes itself when the road clears. What you see is what is wrong now, not a list of everything that ever went wrong.
A vehicle seen an hour later didn't drive slowly — it stopped for lunch. Corridor times use a trimmed median, so a handful of stopped cars can't make a clear road look congested.
No lane closures, no loops in the asphalt, no radar to certify and re-certify. A camera and a drawn line replace the entire installation.
The same cameras and the same server already running MIT Eyes can count traffic — and the same plate reading serves parking and investigations.
Competitors sell four systems: four installs, four vendors, four invoices. MIT sells one platform where every module reuses the cameras, the AI and the data of the others — so each solution you add costs a fraction and returns in full.
See MIT Traffic running on your own cameras — a pilot takes a day, not a quarter.