Crossing Lab

Crossing Lab explains one potential intersection conflict from an elevated view. A real CC0 photograph remains the visual source of truth while a compact, authored motion fixture makes direction, speed, predicted path, pairwise conflict timing, TTC, uncertainty, and simple timing interventions easy to read.

The interface has one sequence:

Watch the three tracks → understand the conflict → test a timing change

Research interface, not a crash detector. The photograph is real. The tracks, velocities, confidence values, TTC, and collision-likelihood score are authored demonstration data. They were not inferred from the single image and must not be used for roadway, vehicle, or safety decisions.

Application walkthrough

Open the full-resolution MP4 · Open the poster frame

The walkthrough media is captured from the running application. It shows the full Watch → Conflict → Test sequence, including turning-vehicle selection, the likelihood uncertainty band, a protected-turn timing counterfactual, the replay action, and both themes.

What makes the example understandable

The screen deliberately separates evidence from explanation:

• The background is a 1920×1280 elevated photograph of a real Vancouver intersection by Ferdinand Stöhr, dedicated under CC0.
• Solid colored paths show authored track history.
• Dotted extensions show each track’s constant-motion fixture forecast.
• An arrow at every actor communicates heading and velocity.
• The shared conflict point is labeled directly on the road surface.
• Both actors list their time-to-conflict and authored arrival time side by side.
• The collision-likelihood percentage always says illustrative, includes a ± uncertainty range, and is followed by a plain-language explanation.
• The Test step changes only the turning vehicle’s arrival timing, so the counterfactual is easy to compare.

The pedestrian is context rather than part of the primary vehicle pair and can be hidden with a single checkbox.

Why these measures are shown

The interaction model follows conventions from primary transportation and motion-forecasting sources:

• FHWA’s Surrogate Safety Assessment Model evaluates conflicts from detailed trajectory records and defines a conflict as a situation in which users are likely to collide without evasive action.
• FHWA’s foundational surrogate-measures report treats conflict points, TTC, post-encroachment time, speed differential, and deceleration as surrogate measures—not substitutes for observed crash data.
• Argoverse 2 motion forecasting represents actor state with time-indexed position, heading, and instantaneous x/y velocity, and explicitly distinguishes focal, scored, and contextual tracks.
• Waymo Occupancy Flow Fields combines occupancy with motion direction; Crossing Lab similarly keeps direction visible instead of presenting a risk heatmap alone.
• NHTSA’s intersection collision-avoidance research uses controlled test procedures for crossing-path scenarios. Because this portfolio example lacks calibrated sensors and test-track ground truth, its percentage is explicitly illustrative rather than presented as calibrated.

The authored fixture

Road users

Track	Role	Recorded fixture speed	Confidence
V-21	southeast through vehicle	12.6 m/s · 45 km/h	96%
V-08	northbound turning vehicle	8.4 m/s · 30 km/h	93%
P-04	crosswalk context pedestrian	1.4 m/s · 5 km/h	88%

Confidence here is fixture metadata used to demonstrate a review interface. It is not a perception-model result.

Conflict calculation

For the recorded-flow fixture:

V-21 conflict arrival = 4.25 s
V-08 conflict arrival = 4.48 s
arrival gap           = 0.23 s

At the default review time of 2.45 seconds, TTC is 1.80 seconds for V-21 and 2.03 seconds for V-08. These are remaining times along the authored image-plane tracks, not photogrammetric measurements.

The demonstration likelihood uses an intentionally small, inspectable formula:

illustrative likelihood = round(52 × exp(-0.9 × arrival_gap))

The UI attaches an authored uncertainty band of ±9 percentage points to the observed fixture, ±6 to early braking, and ±3 to the protected-turn example. The formula exists to make UI states deterministic; it has not been fitted, calibrated, or validated against crash outcomes.

Timing comparisons

Control	Turning-vehicle arrival	Purpose
Observed flow	4.48 s	baseline authored encounter
Early brake	5.62 s	show how earlier deceleration separates arrivals
Protected turn	7.10 s	show a separate phase after the through path clears

Selecting a control immediately updates the trajectory, turning-vehicle speed, TTC, gap, likelihood, uncertainty, and plain-language result. “Replay this timing” restarts the complete eight-second fixture with the selected control.

Interaction inventory

Every visible control changes an observable state:

Control	Result
Watch / Conflict / Test	switches the explanation and sets a useful playhead state
Light / Dark	persists the selected theme locally
Pedestrian context	shows or hides P-04
Actor on image or in list	changes the velocity and heading readout
Play / Pause	starts or stops the fixture clock
Timeline	seeks to an exact fixture time and pauses
Restart	resets to 0 seconds and plays
Observed / Early brake / Protected turn	reruns the deterministic timing comparison
Test a safer timing	advances from the conflict explanation to Test
Replay this timing	restarts the selected comparison

Keyboard focus, Enter/Space activation on SVG actors, status announcements, responsive layouts, reduced-motion behavior, and light/dark palettes are included.

Architecture

flowchart LR
    A["CC0 overhead photograph"] --> V["Intersection view"]
    B["Authored time-indexed tracks"] --> C["Interpolation"]
    C --> D["Heading + velocity vectors"]
    C --> E["Forecast extensions"]
    C --> F["Pairwise conflict point"]
    G["Timing intervention"] --> C
    F --> H["TTC + arrival gap"]
    H --> I["Illustrative likelihood + uncertainty"]
    D --> V
    E --> V
    I --> P["Plain-language review panel"]

Loading

Layer	Responsibility
app/components/MotionLab.tsx	three-step workflow, clock, controls, themes, SVG interaction, accessible status
app/lib/overhead-scenario.ts	source provenance, actors, interpolation, heading, forecasts, TTC, deterministic comparison
public/scenarios/vancouver-overhead.jpg	bundled 1920×1280 CC0 source photograph
tests/overhead-scenario.test.mjs	image integrity, pair timing, intervention ordering, velocity and path contracts
backend/app/	optional typed FastAPI research fixtures retained for extension work

Run locally

Prerequisites: Node.js 22.13+ and Python 3.11+.

npm install
python3 -m venv backend/.venv
backend/.venv/bin/python -m pip install -r backend/requirements.txt
npm run dev

The overhead application is fully local and does not require the Python API. The API can still be started separately for research extensions:

npm run api

Verification

npm run verify

The gate covers the production build, strict TypeScript, ESLint, server-rendered HTML, source-media dimensions, actor taxonomy, shared conflict geometry, TTC/arrival values, intervention ordering, heading and forecast output, WCAG AA palette contrast, DOM-level interaction of every workflow control, the legacy deterministic browser/API parity layer, FastAPI schemas, and backend tests.

Accuracy boundaries

• The background is a single elevated photograph, not a calibrated overhead camera feed or orthorectified map.
• SVG positions are image coordinates. Distance is not measured in meters.
• Tracks, boxes, velocity, heading, confidence, arrival time, TTC, risk, and uncertainty are authored fixtures.
• The likelihood is not statistically calibrated and is not a probability of a crash in the photographed scene.
• The intervention comparison changes an authored arrival schedule; it does not simulate braking dynamics or a traffic controller.
• No detection, tracking, forecasting, planning, warning, or actuation model is deployed.
• Production use requires calibrated multi-frame sensing, evaluated tracking, map alignment, probability calibration, rare-event validation, safety engineering, and applicable regulatory review.

Attribution and derivative details are recorded in THIRD_PARTY_NOTICES.md.

No source-code license has been selected; copyright remains with the repository owner. Third-party media remains under the terms recorded in the notice file.