Skip to content

Browser simulator

The browser simulator runs the decision pipeline on the scenario suites in a bird's eye view, one frame at a time, with no install. This page describes its controls and how its numbers relate to the offline testbench.

Open the simulator. From a checkout, open simulator/index.html in any browser; it is a single self-contained HTML file, and simulator/scenarios.extended.js next to it adds the suites beyond conformance.

Panels

Panel Contents
Bird's eye view Road, crosswalk, cameras with their field of view, agents and their tracks, in the scenario frame
Decision pipeline A flowchart of the stages (detection, classification and ground projection, tracking, then the three decision checks) with the active path lit and the current output: idle blue, safe green, warning orange, alert red blink
Timeline Scenario time, duration, play, reset, loop and a speed slider
Telemetry Number of objects, number of tracks and the minimum bike-pedestrian distance
Systematic error budget The localisation error that the chosen camera setup and uncertainty parameters introduce for the scenario being played

Suites and scenarios

The suite bar along the top switches between suites. Conformance holds the paper's 24 scenarios and is always present. Further suites (extended, stress) appear when scenarios.extended.js is present next to the page; without it the page runs with the conformance suite alone.

Selecting a scenario writes its slug into the URL fragment, so a link such as simulator/#fast-approach opens that scenario directly. Other pages of these docs use this to link to individual scenarios.

Pipeline variants

The pipeline selector offers seven configurations. The first is the deployed configuration; the last three are the decision-rule baselines from the paper, run with the same front end.

Variant Memory (frames) d_min (m) d_max (m) Min displacement (m) Lookback (frames) Speed-adaptive Rule
Selected 58 1.9 24.8 0.147 2 yes pairwise
Conservative 60 1.5 15.0 0.05 3 no pairwise
Speed-Adaptive 60 1.0 20.0 0.05 2 yes pairwise
Proximity Only 0 2.0 10.0 0.1 3 no pairwise
Distance Only 58 1.9 10.0 0.0 2 no distance_only
Naive Closing 58 1.9 24.8 0.147 2 yes naive_closing
TTC 58 1.9 24.8 0.0 2 yes ttc, threshold 3 s

The values are written into the page and mirror decision_testbench.PIPELINE_CONFIGS. Decision pipeline explains each parameter and rule.

Cameras

The setup buttons place cameras in four presets: single pole, cross-cam (a second pole at (5, 13) facing the first across the crosswalk), centre and overhead. The + and − buttons add and remove cameras, and the sliders set the selected camera's position (X, Y), height, field of view and heading. The default camera stands at (5, −13) m, 5.5 m high, with a 220° field of view facing the crosswalk.

Uncertainty parameters

The uncertainty panel sets the detector and tracker model that turns ground truth into noisy observations:

Group Parameters
YOLO detection Average recall and precision for small, medium and large objects (COCO size buckets), latency, high and low confidence thresholds
BEV projection Ground tilt uncertainty (the street is not perfectly level)
ByteTrack Kalman filter Position and velocity noise, frames before a lost track is dropped, IoU threshold, frame rate

Changing them changes how often agents drop out and how far their projected positions wander, which is the quickest way to see why the decision stages need memory and a lookback.

Relation to the testbench

The simulator is a JavaScript port written for inspection. The numbers in the paper and on these pages come from the Python offline testbench, which shares its code with the deployed system. Use the simulator to understand a scenario and the testbench to measure one.