RIA Labs exhibits at its own stand
Schematic scene · vector drawing in metres, not the camera output
RIA Smartcity reads the video a city already has and returns measurements and facts, not recordings: how many go past and of what kind, how fast they really come, where they nearly crash, who walks and where they want to cross, how much of the kerb is taken, whether the street flooded or a street light went out. Every vertical on the same image.
“Under almost every actuated traffic signal in the world there is a wire loop buried in the road. It answers one single question: is there something on top of me? Yes. No. Yes.”
One bit per vehicle, once. Not what it is, not how fast it is coming, not where it is going, not when it arrives. Four fields the loop leaves empty.
“The same question, answered by a camera with tracking. And for every vehicle, on top of that: what it is, how fast it is coming, where it is going, and in how many seconds it reaches the stop line. It is not a different sensor: it is a strict superset of the one already buried.”
And what is detected is not only a vehicle: it is everything that moves —cars, buses, lorries, motorcycles, bicycles and people—. That is where the five verticals begin, in the same frame.
YOLOX-s · ONNX Runtime · CUDA · 19.5 ms per frame · measured on the box, today“An uncalibrated camera measures in pixels, and a pixel is not a unit of management. With two sections and two real distances, the scene enters a metric plane: that is where metres per second, the gaps between vehicles and the queue come from.”
The projection you are watching rotate is the same operation the product performs: the homography that takes the camera view to the plane of the carriageway. That plane is used by all five verticals: without metres there is no speed, no near miss, no kerb occupancy.
“And where calibration does not reach, the system says so.
This is not estimated, it is flagged.”
Everything that moves leaves its track. Hundreds of curves later, the junction shows its desire lines, and out of the cloud come the three figures a traffic engineer uses every day: flow, v85 and headway.
But the track is not a traffic figure, it is a figure about the scene: the same curves, filtered by class, are where people want to cross —not where the crossing happens to be painted—, how long the person waiting for the bus stands there and how many kerbside bays are taken at each hour, against a declared capacity.
“None of these curves has a name. They are shapes, class and hour. The video is processed on the box and discarded.”
Two controllers run in parallel over the same traffic: the fixed-time one that is installed and one actuated by time of arrival. Two queues, side by side, and the cumulative delay pulling apart.
“This is not a new control theory. It is the canonical doctrine, the one in the FHWA manual and in the 2.5 to 5 second dilemma zone, implemented with a better sensor.”
27–28% delay reduction · measured on a development clip · the figure only appears with 15 vehicles and 2 cycles · saturation flow 1,800 veh/h/lane“You don't have to change the traffic lights:
you have to change the sensor.”
“A crash is a figure that arrives late and one at a time. A near miss is the same figure, early and in volume. Bellevue spent ten thousand dollars on paint and signage and cut its critical conflicts by sixty per cent.”
And the method does not tell vehicles from people apart: it measures the crossing of any two trajectories, so the same instrument counts car against car, car against pedestrian and car against bicycle. When one of the two is a vulnerable road user —pedestrian, bicycle or motorcycle— the same number of seconds weighs more, and the system says severity went up because of that.
Bellevue · USD 10,000 in paint and signage · 60% fewer critical conflicts · the only risk colour in the whole film“When a resident demands “put a traffic light here”, the council answers with data.”
“Everything you have seen so far runs today, on one camera, on a box the size of a modem. What follows is not another technology: it is the same one, repeated.”
And the layers do not end at traffic either. On this same image live five verticals: mobility, road safety, public space, parking and environment. None of the four that follow the first asks for another camera or another civil works contract: they ask for a line, a zone or a check.
And here is the limit, said before anyone asks: one camera measures one approach. Corridor coordination, actuation on the installed controller, cross-street demand —simulated today—, turning movement counts by movement and phase, and multi-camera scale in production are designed and documented, not running. Below, each one carries its stamp.
“The same camera, the same image, and the question changes: not how many cars go past, but how many people walk and where they want to cross.”
Pedestrian and bicycle are classes of their own from the first frame. A counting line does not return a total: it returns the breakdown by class and by direction. And the same anonymous tracks, filtered by class, are the desire lines of the people: you can see the cloud does not land on the paint.
Pedestrian and bicycle classes · counting line with ab and ba directions · anonymous track of at most 12 points, with no identity and no fine timestampThe junction is not only a place people pass through: it is a place where people wait. A zone with a dwell role measures in seconds how long each person stays, and publishes the mean and the 85th percentile. When there are too many at once and it is sustained, the crowding alert opens.
Dwell-role zone · stay in seconds, mean and p85 · default threshold 15 people sustained 60 s, lowered to 6 in this zone“The kerb is the most fought-over surface in the city, and the one nobody measures.”
Zones with a parking role on the same plane: how many are there, against a declared capacity, in vehicles and in per cent; how many arrived and how many left; and how long each stay lasted.
10.0 s still to count as a bay · capacity declared by hand, never inferred from the geometry · the linear metres of kerb are NOT measuredDouble parking is not a separate detector: it is a vehicle standing still inside a carriageway-role zone, below 3 km/h, for 45 seconds, with double confirmation. The same episode engine that traffic uses.
The loading bay as a role of its own is NEXT. What can already be done today is to declare that zone with a timetable, and have it change role according to the hour.
“Everything you have seen so far is tracking: frame by frame, each thing with its identity. What follows tracks nothing: it looks.”
One check per zone every half hour, on a crop of the scene and not on the whole frame. And three possible answers, not two: positive, negative and undetermined.
Is there fly-tipping? Did the street flood? Did the street light go out? The first two —and smoke— are answered by a vision language model on the crop. The street light is not: it is answered with the local brightness of the crop against a reference calibrated on site.
And when the carriageway cannot be seen, the flooding check answers undetermined: there is no false negative hidden inside a “no”.
1 check every 30 min per zone, floor of 5 min · crop of at most 768 px per side, discarded after the query · minimum confidence 0.70 · two consecutive checks to open the episodeCity sensors,
for intelligent decisions.
Keep scrolling: the thesis and the board
The scene is a four-corner junction drawn in world coordinates, in metres: carriageways, four corner splays, pedestrian crossings, kerbs, stop lines, traffic signals and building footprints. There is no photograph anywhere. The clock stays frozen at 08:14:07 and the traffic moves in a continuous loop.
Layer 1, the road: a rectangle of induction loop buried under the lane, switching on and off once per vehicle, and four empty fields: class, speed, direction and time of arrival.
Layer 2, the camera: the oblique view of the junction, detection boxes over the moving vehicles and the four fields filled in. Label: YOLOX-s, ONNX Runtime, CUDA, 19.5 milliseconds per frame, measured on the box, today.
Layer 3, the metric: the projection rotates from the camera view to the plan view. Two calibration sections appear, the declared separation between them and the carriageway width as engineering drawing dimensions. The dimensions belong to the drawn schematic: avenue carriageway 14.00 m, cross street 11.00 m, pavement 4.00 m, corner splay 5.00 m, separation between sections 14.00 m. Inside the calibrated band the drawing stays sharp; outside the band it degrades to grey, because there the system does not report: it flags.
Layer 4, time: everything that moves leaves its track and hundreds of accumulated tracks form the desire lines of the junction. Three figures are distilled —flow, v85 and headway between vehicles— computed on the drawn scene, not measured in the field, and only from a sample of 15 vehicles. The same tracks, filtered by class, are the public space instrument: where people really cross, how long they stay at the stop and how much kerb is taken at each hour.
Layer 5, the decision: the traffic signals appear along with a detail showing the two queues, fixed-time and actuated by time of arrival, over the same arrivals; below, the two cumulative delay curves pulling apart, and the 2.5 to 5.0 second dilemma zone marked over the approach. The 27–28% delay reduction measured on a development clip only appears once the sample reaches 15 vehicles and 2 cycles.
Layer 6, risk: the scene empties and two road users are left almost touching. The post-encroachment time, PET, is drawn between the instant one leaves the conflict point and the instant the other enters it. In the margin, the day's near misses ordered by severity, by PET and by TTC. The method measures the crossing of any two trajectories, so it covers equally a conflict between vehicles and one between a vehicle and a pedestrian or a cyclist; with a vulnerable road user severity goes up, and the system declares that it went up because of that. It is the only place in the film where the risk colour appears, and the day's values are not published here because the scene is a schematic.
Layer 7, what is not in the image: the view pulls back and the junction becomes one point among other points on a schematic map. The measured junction is labelled “1 camera, 1 approach, 5 verticals”: mobility, road safety, public space, parking and environment on the same frame. The neighbouring points are marked as next: corridor coordination, turning movement counts by movement and phase, cross-street demand, actuation on the installed controller and multi-camera scale.
Layer 8, public space: the view returns to the south-west corner and the subject becomes the pavement. Three counting lines are drawn —the avenue crossing, the street crossing and the south pavement— and each publishes its total by direction; the product also breaks it down by class, and pedestrian and bicycle are classes of their own. Over the crossings the desire lines of the people accumulate, anonymous tracks of nine vertices that do not land on the paint: it counts how many cross off the crossing. On the corner splay there is a zone with a dwell role where the stay is measured in seconds, mean and 85th percentile, and the crowding alert lights up when there are more than six people at once sustained for sixty seconds; the product default threshold is fifteen and this zone lowers it. All the figures are computed on the drawn scene.
Layer 9, the kerb: the view drops to the south approach and three polygon zones appear over the kerb, two with a parking role and capacity declared by hand, and a third marked as loading with a NEXT stamp, because the loading bay as a role of its own is not built: what exists today is a zone that changes role by timetable. Occupancy is reported in vehicles and in per cent against that capacity; the linear metres of kerb are neither measured nor inferred, and that is declared. Arrivals, departures and mean stay are kept. A vehicle stands still in the lane inside a carriageway-role zone: below three kilometres per hour, for forty-five seconds and with double confirmation, that is double parking, and the ones coming down swerve around it.
Layer 10, the state of the scene: the projection returns to the camera view and everything outside a watched polygon goes dark. Inside remain the kerb corner, a gully and a street light. This is the product's other engine: there is no frame-by-frame tracking, there is one check per zone every thirty minutes, with a floor of five, on a crop of at most seven hundred and sixty-eight pixels per side that is discarded after the query. Four checks: fly-tipping, flooding, smoke and street light out. The first three are answered by a vision language model; the street light is answered with the local brightness of the crop, the average of the maximum of the three channels, against a reference calibrated on site, below fifty per cent. Each check has three possible answers, positive, negative and undetermined, and flooding turns to undetermined when the carriageway cannot be seen. Two consecutive checks are needed to open an episode.
Close: the world withdraws, the clean drawing is left and the sign-off: city sensors, for intelligent decisions.
01 · The thesis
A camera over a junction is not a traffic instrument. It is an instrument, and traffic is merely the first question we asked it. The product's other four verticals —road safety, public space, parking and environment— do not ask for another camera, or another pole, or another civil works contract: they ask for another reading of the same pixels. That is the product's commercial argument, and it is also its technical one.
What the next vertical costs
What is not paid for twice
And where it is not free, we say so: the pavement has to be in the frame to count pedestrians, the approach needs resolution to sample helmets, the street light needs a normal night as a reference, the scene checks have a daily budget, and speed does not come out without two distances measured in the street. A new vertical is a configuration, not a miracle.
That is the arithmetic that closes the deal: the first vertical pays for the installation and the second one costs a configuration. The film above is that argument told as form — the clock does not move because the marginal cost of the next layer is zero.
02 · The board
Here the site stops being linear. There are five verticals and, under each one, real questions of municipal management: each opens a card with what is measured, what decision it enables and what it takes to have it. Every card carries its stamp, and the stamp is not negotiable. What is designed and not yet running appears here by name, not in the small print.
V1 · Vertical
Asked for by Traffic and Urban Mobility
Used by whoever retimes the corridor and whoever signs off the works
What moves through here, of what kind, at what speed, in which direction and with how much queue. It is the product's most mature vertical and the only one with a measured figure of its own, so it is the one that opens the door in most cities. It is also the one most often mistaken for the whole product: it is one of five.
Third-party validation: VivaCity measures 97% classification accuracy, validated by Transport for London.
Measured in development: 27–28% delay reduction against fixed-time, on a development clip, with a minimum sample of 15 vehicles and 2 cycles before the figure is shown.
V2 · Vertical
Asked for by Traffic, and by the mayor when the neighbourhood complains
Used by whoever decides where the paint goes
A crash is a figure that arrives late and one at a time. A near miss is the same figure, early and in volume: with weeks of operation, the city's risk map before the first case file. It is the product's differentiator and the vertical with the most eloquent international case: it is fixed with paint, not with civil works.
Third-party validation: Bellevue cut critical conflicts by 60% with USD 10,000 of paint and signage. Wavetronix reports crash reductions of 31 to 50% for angle crashes and 75% for rear-end crashes.
Third-party validation: Bellevue, 60% fewer critical conflicts with USD 10,000 of paint and signage. A third party's result, not this product's.
V3 · Vertical
Asked for by Public Space and Works
Used by whoever draws the crossing, the cycle lane and the square
The same camera that counts cars counts people and bicycles, and their tracks are the product's most eloquent instrument: where people want to cross, not where the crossing happens to be painted. Sustainable mobility areas have a political agenda and zero data: this is the vertical that gives them some.
V4 · Vertical
Asked for by Controlled Parking, Traffic and Enforcement
Used by whoever sets the tariff and whoever sends out the inspector
The kerb is municipal revenue and is usually managed blind; double parking is the resident's number one complaint and on top of that it strangles the traffic and the buses. Both are measured with the same instrument: a zone, a stillness threshold and a clock. Without reading a single number plate.
V5 · Vertical
Asked for by Environment and Public Services
Used by whoever dispatches the crew and whoever signs the sustainability report
Here the sensor changes engine: instead of tracking things that move, it answers questions about the state of the scene, every half hour, on a crop of the zone being watched. Fly-tipping does not appear between two frames. It is the vertical that often opens the door when the city's pain is not traffic.
V0 · Cross-cutting
What is coming, with its name and its scope
It is designed and documented, and it is what comes next. It appears on this page with its name and its scope because it is part of the product that is coming: each of these points already has a defined what it measures and what decision it enables.
Real actuation on the installed controller
Dry contact or NTCIP 1202. The comparator and the actuated logic run; closing the loop against the junction's controller does not.
NextCorridor coordination
The green wave between junctions. Today each camera solves its own approach, and the network optimum is another league.
NextCross-street demand
Today it is simulated as a queue, and it is declared that way on screen. Measuring it for real means putting up the second camera.
NextMulti-camera scale in production
One junction measuring for real, today. The network of junctions is the same technology repeated, and it is not repeated yet.
NextTurning movements by movement and phase
The canonical TMC, with U-turns and prohibited manoeuvres. What runs is the origin-destination matrix by pairs of zones, and that is not the same thing: we point it out ourselves because in this sector the name gets used loosely.
NextGeometric double-parking detector
Today double parking is read from the vehicle stopped on the carriageway events. The dedicated detector that tells it apart from any other stop is missing.
NextHourly baseline for congestion and crowding
Both thresholds are absolute. “Is this a lot for this time of day?” needs weeks of data: with two days, any baseline is noise dressed up as signal.
NextPeople heat map
The heat grid exists and today it filters vehicles. Making it about people is changing one filter — and until it is changed, it is not offered.
Next03 · Integration
The junction's data belongs to the council, not to the supplier. Four ways out —the same for all five verticals, because there is a single interface— and each with its stamp: three run today, one is the next stage.
Our own REST interface
Counts, speeds, turns, dwell, parking, helmets, emissions, conflicts, events and fleet health over HTTP. The same interface the live analyser uses: there is no shop-window API and a real one. A malformed request returns an explicit error with its code, never a silent truncation.
Running todayCSV in one click
For whoever works in a spreadsheet, which is almost everyone: semicolon separator and decimal comma, which is what opens properly in a European Excel. Every export travels with its sample and its time window.
Running todayNGSI-LD from the city data model
TrafficFlowObserved entities from the transport Smart Data Model, to enter an already installed urban platform without writing an adapter. With the limitation declared: that model's vocabulary has no value for pedestrian, so the pedestrian falls outside that serialisation and stays available in our own interface. The pedestrian flow model is another entity and it is not implemented.
Running todayDry contact or NTCIP 1202
The two standard ways into the controller that is already installed: emulating the loop's call into its detector terminal block, or speaking NTCIP if the equipment is modern. There is no need to change the traffic light. Designed and documented; not running yet.
NextAnd what is not a way out but a property: the data lives on the box. The per-minute aggregates, the events and the summarised tracks are kept in a local database with its declared retention, so the junction keeps measuring with the link down and no decision depends on a cloud. The video, by contrast, is not stored anywhere.
04 · Provenance
There is not one figure on this site that is not in this table. If a figure has no provenance, it is not published.
19.5 to 22.9 ms per frame · 12 frames per second
Measured on the box, today, with CUDA, on a live calibrated source. The figure shown in the film is 19.5 ms. The milliseconds that appear in the deployment log of 13 August are from the previous installation without a GPU and are not used anywhere on this site.
27–28% delay reduction
Measured in development, against fixed-time, on a development clip, with a minimum sample of 15 vehicles and 2 cycles before the figure is shown.
Zero false positives on wet asphalt
Measured in development, on the scene checks.
1,800 veh/h/lane · dilemma zone 2.5 to 5.0 s · minimum green · optimum cycle
Doctrine, with its source: a prudent saturation flow against the ideal of 1,900; dilemma zone and extension formula from the Wavetronix document; minimum green and clearance tables from the FHWA detector manual; two objective modes from TRL's MOVA; optimum cycle from Webster.
97% classification accuracy · 23% less journey time
VivaCity: accuracy validated by Transport for London; the journey time reduction is from Salford, over 15 junctions and 4 years in Greater Manchester. A third party's result, not this product's.
60% fewer critical conflicts · USD 10,000
Bellevue: paint and signage. A third party's result.
31 to 50% for angle · 75% for rear-end
Wavetronix: crash reductions reported in their document. A third party's result.
COPERT v5.9 · EMEP/EEA · 4 speed bands
A declared model, not a measurement. The emission factors were computed from the official COPERT v5.9 coefficients of the European EMEP/EEA inventory and weighted for the Argentine fleet, in four speed bands. The table is editable and versioned: a council with data on its own fleet replaces it. Declared scope: hot emissions, without particulate matter and without an instantaneous acceleration model. There is no gas sensor at the junction.
Product thresholds: 45 s · 10 s · 15 people and 60 s · 12 points · 15 s bands · factor 1.5 · check every 30 min
Installation parameters, not measurement results. These are the values the product is installed with and each is tuned per site at commissioning: stillness to count as stopped on the carriageway (45 s) and as occupying a bay (10 s), the crowding default (15 people sustained 60 s, which each zone can override), the hard cap on points per anonymous track (12), the aggregation band for dwell (15 s), the severity threshold multiplier with a vulnerable road user (1.5) and the default frequency of a scene check (30 min, with a hard floor of 5 min that is not configurable).
Drawing dimensions: 14.00 · 11.00 · 16.00 · 4.00 · 5.00 · 1.80 · 9.00 m
Avenue carriageway, cross street, separation between the two calibration sections, pavement, corner splay, side of the induction loop and camera height. They are dimensions of the schematic drawn in the film, not of a real site: the scene is a typical junction in world coordinates. They are there to show what an engineering drawing dimension looks like, not to declare the geometry of a real junction.
05 · The proof
Everything the film told you can be checked against the junction that is measuring at this moment: the live analyser, with its calibrated source, its counts and its conflicts. It is not a mock-up.
Access is by invitation: the live junction asks for a username and password.
Video analytics applied to the public highway
RIA Smartcity · 2026