banshee-data/sf-street-speeds

San Francisco street-speed dataset

Measure velocity, not identity

This dataset publishes replayable LiDAR observations from stationary roadside surveys in San Francisco. Raw PCAPNG captures are kept beside derived browser-scene exports, with a machine-readable manifest joining the two.

The corpus is intended for reproducible research in roadside perception, LiDAR processing, object tracking, trajectory and speed estimation, traffic behaviour, and static-scene reconstruction. Reviewed reference labels cover one capture so far (see Labels). It does not provide ground-truth tracks, trajectories, or reference speeds.

There are no camera images, licence-plate photographs, or intentionally identifying fields. There are raw LiDAR packets. They are evidence, not decoration, and should be handled with the care due to a detailed measurement of a public place.

                                ░░░░░░                                  ░░░░░░
   ░░░░░    ░░░░░░░░░░░    ░░░░░ :::. ░░░░░░░░░░    ░░░░░░░░░░░    ░░░░░ :::.
  ░▒▒▒▒▒░░░░▒▒▒▒▒▒▒▒▒▒▒░░░░▒▒▒▒▒ .:.  ▒▒▒▒▒▒▒▒▒▒░░░░▒▒▒▒▒▒▒▒▒▒▒░░░░▒▒▒▒▒ .:.
░░▒▓▓▓▓▓▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒ .::.::: : ▓▓▓▓▓▓▓▓▓▓▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒ .::.:.: :
▒▒▓.  . ▓▓▓▓████ ... . ▓▓▓▓ :....:::::. ▓▓▓.  . ▓▓▓▓████ ... . ▓▓▓▓ :.::..::::
▓▓█ ..::████████ :.::: ████ :: :  : . ::▓▓▓ ..::████████ :.::: ████ :: :: : ::
:.: : ::████:.:: : ::: ████ .: : .  ... :.: ::::████:.:: : ::: ████ .::: .:::.
.::  .  .   ::.:    .  .          .     .::  .  .   ::.: ::::  .          .
    .     .       .
                         ▄▀▀▀▀▀▀▄
                       ▄▄█▄▄▄  ▄▄▌
                         ▌▐▀▀▀█  ▌
▀▀▀▀▀▀█████▀▀▀▀▀▀▀▀▀█▀▀█▀▐▀  ▄▄▀▀█▀▀▀▀▀▀▀█▀▀██▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀
▄▄▄▀▀▀░░ ░ ░ ░▒▓▄▄▀ ▄▄▀▀ ▄▓▀▀▀  ░▐▌       ▀▀▄▄▀▀▄▄▓▒ ▒   ░ ░    ░  ░   ░  ░
 ░ ░░ ░ ░ ▒ ▄▄▀▀▄▄▀▀▄▄▄▄▓░   ░▄  ░█           ▀▀▄▄▀▀▄▄  ▒ ░   ░   ░   ░ ░    ░
  ░  ░ ▒▄▄▀▀▄▄▀▀   ▀▄▀█░ ▄▄▀█▀░░ ░█               ▀▀▄▄▀▀▄▄   ▒ ░    ░      ░
░░ ▒▄▄▀▀▄▄▀▀       █▀▀▓▀▀▌  ▐▌░ ░░▒▌                  ▀▀▄▄▀▀▄▄   ░ ░   ░  ░  ░
▄▄▀▀▄▄▀▀           ▓  ▓  ▌  █▀▀▀▀▀▀█                      ▀▀▄▄▀▀▄▄  ▒░ ▒
▄▄▀▀               ▓▀▀▓▀▀▌ ▐▌░  ░  ▐▌                         ▀▀▄▄▀▀▄▄   ░ ▒
                   ▓__▓__▌ █░  █▄░ ▐▌                             ▀▀▄▄▀▀▄▄   ░
                   ▓  ▓  ▌ █░ ▐▌ █░ █                                 ▀▀▄▄▀▀▄▄
▄▄▄  ▄▄▄▄▄▄  ▄▄▄▄  ▓  ▓  ▌▒▒▒▒▒  ▒▒▒▒   ▄▄▄▄▄  ▄▄▄▄   ▄▄▄▄▄  ▄▄▄▄▄  ▄▄▄▄▄▄▄  ▄
█▀ ▄█████▀ ▄████▀  ▄████   ████   ████  ██████  ▀███▄  ▀████▄ ▀████▄ ▀██████▄
 ▄█████▀  ▄████▀  ▄████▀  ▄████   █████  ██████  ▀███▄  ▀████▄  █████▄ ▀██████
█████▀  ▄█████   ▄████▀   █████   █████   ██████   ████▄  █████  ▀█████▄ ▀████
████▀  ▄█████   ▄█████    ████▀   ▀████▄   ██████   ████▄  ▀████▄  ▀█████  ▀██
██▀  ▄█████▀   ▄█████    ▄████     █████    ██████   ▀████▄ ▀█████  ▀█████▄  ▀
▀   ▄█████▀   ▄█████▀    █████     ██████    ██████   ▀████▄  █████▄  ▀█████▄

At a glance

The dataset is a growing collection of stationary roadside surveys rather than a fixed benchmark release. New captures and derived products may be added as field coverage expands.

Each PCAPNG covers an operator-approved interval in which the sensor platform was parked. static describes the platform, not the street: vehicles, pedestrians, cyclists, and the ordinary business of a city continue moving through the point cloud.

The source recorder may roll files during a survey. Each published PCAPNG clips the contributing source files to one indexed site interval and merges the pieces in capture order. The manifest retains source names, byte counts, and SHA-256 digests so that this lineage is inspectable rather than merely asserted.

Quick start

manifest.json is the authoritative index. Start there rather than walking the directory tree.

Inspect the available captures:

jq '.[] | {
  capture_id,
  timestamp,
  duration_seconds,
  s2_l13,
  site_slug,
  raw_path,
  scene_path
}' manifest.json

Select a site:

jq '.[] | select(.site_slug == "columbus-broadway")' manifest.json

Paths in the manifest are relative to the dataset root, so analysis code can use the manifest to discover raw and derived artefacts without reconstructing paths from filenames.

Research uses

The corpus is suitable for experimentation and evaluation involving:

• LiDAR packet and point-cloud processing
• background and static-scene modelling
• clustering and object tracking
• trajectory reconstruction
• speed and motion estimation
• traffic-behaviour measurement
• roadside perception pipelines
• reproducibility and pipeline-comparison studies

It is not a supervised perception benchmark. One capture carries reviewed reference labels (Labels); the rest carry none, and the corpus contains no ground-truth tracks, trajectories, or reference speeds. Derived products should not be treated as ground truth merely because they were generated by the project pipeline.

Raw and derived representations

Representation	Role
manifest.json	Authoritative corpus index joining captures, locations, provenance, and derived artefacts
raw/lidar/*.pcapng	Replayable raw LiDAR evidence for a capture interval
derived/scenes/*_vrlog/	Chunked browser-scene representation derived from the corresponding raw capture
labels/<stem>/<set>/	A label set: reference objects and their per-frame membership, citing the raw capture
labels/<stem>/<pack>/	A label set's view: the run the labels were drawn on, which the annotation tools open

The raw capture is the durable research record. Derived scenes are convenient views intended to make the same observations easier to inspect and process. Derived representations may evolve as the pipeline improves.

The scene exports are directories because the browser format is chunked. Each contains its manifest and one or more parts with headers, timelines, indexes, and compressed frame chunks. Repacking that tree as a pretend single VRLOG would make the name simpler and the data less useful, which is the wrong exchange rate.

The historical provenance available for this corpus does not establish every sensor-level detail needed to describe a universal decoded point schema. Where sensor model, processing build, or other metadata is unknown, the manifest records null rather than inferring a value.

Layout

README.md
manifest.json
raw/
  lidar/
  radar/
derived/
  scenes/
  tracks/
  maps/
  metrics/
labels/
  <capture stem>/
reports/

Directories describe the kind of artefact. Dates, S2 cells, and sites remain in filenames and metadata rather than creating another hierarchy of folders.

The current corpus populates raw/lidar/, derived/scenes/ and, for one capture, labels/. The other directories reserve stable homes for later radar, track, map, metric, and report artefacts. Their presence is not a claim that those artefacts exist.

Labels

Reference labels say which returns belong to which road user, frame by frame. They are published as label sets, beside the capture they describe:

labels/<capture stem>/
  <label set name>/   labels.json and the head files: objects, windows, masks
  <pack directory>/   its view: one run's points of the capture, with physical
                      references, linked to the label set beside it

A label set cites returns by a digest of the sensor's UDP payload, so it applies to the raw PCAPNG and to any other file holding the same packets. Each mask names the foreground extraction it was drawn on. A return outside that foreground was never shown to the labeller and is unlabelled, not a negative. Masks are partial: they claim what the foreground showed, not the whole vehicle. The view is the run the labels were drawn on, and it is what the annotation tools open. A capture's manifest row lists its label sets in label_set_paths and its views in annotation_pack_paths.

Capture	Label set	Objects	Masks (reviewed)	Physical references
20251205T1417_8085875c_kirkham-9th	kirk0-reference	30	4,102 (4,102)	3 vehicles: 3 bodies and 22 of 44 keyframes reviewed, 2 following decisions

The label sets are maintained with the velocity.report annotation tools. To propose a correction, clone this dataset, make a branch, edit with the project's annotation window, run its repin and check-view checks, and open a pull request. The steps, and what a change commits, are in the project's label set guide.

Filenames

Capture-associated artefacts share this stem:

<timestamp>_<s2-l13-display>_<site-slug>

For example:

raw/lidar/20260902T1320_80858-0f4_columbus-broadway.pcapng
derived/scenes/20260902T1320_80858-0f4_columbus-broadway_vrlog/

• timestamp is the indexed capture start in YYYYMMDDTHHMM form.
• s2-l13-display is velocity.report's human-readable 5+3 form.
• site-slug is lowercase ASCII with words separated by hyphens.

The timezone is fixed in the manifest rather than repeated in every filename. The slug is a label for people, not an authoritative location identifier.

Geographic identity

The filename uses a readable S2 L13 value such as 80858-0f4. The manifest stores its canonical form, 808580f4, and the canonical L10 parent separately. Software should use those manifest fields for grouping and joins.

S2 identifies a geographic partition. It does not uniquely identify a capture, survey, sensor deployment, road, or semantic site. Multiple captures may share an L13 cell without becoming the same observation by a feat of cartographic optimism.

Latitude and longitude come from an operator-marked field map. They are useful neighbourhood-level positions, not surveyed junction coordinates. Do not use them for lane-level registration, legal boundaries, or any task where a few hundred metres would be an exciting surprise.

The manifest is the index

manifest.json contains one object per capture and joins raw evidence to derived artefacts. Treat it as the machine-readable interface to the corpus.

Field	Meaning
capture_id	Stable corpus identity using the minute, canonical L13 token, and site slug
timestamp, timezone	Exact indexed start and its named timezone
filename_timestamp	Minute-resolution timestamp used in filenames
latitude, longitude	Approximate WGS84 field-map position
s2_l10, s2_l13	Canonical S2 tokens for machine use
site_slug	Human-readable site label
sensor_type, sensor_model	Sensor class and model when known
capture_type	How the capture interval was selected
duration_seconds	Exact indexed interval duration
raw_path, scene_path	Dataset-relative artefact paths
tracks_path, map_path, metrics_path, report_path	Related products, or null when absent
pipeline_version, pipeline_git_sha	Splitter build provenance when known
raw_sha256, raw_bytes	Integrity metadata for the published PCAPNG
export_provenance	Static-export policy, operator joins, and contributing source files

Paths use forward slashes and are relative to the dataset root. Unknown values are null. A null is an honest gap; a plausible-looking guess is merely a bug wearing a tie.

Verify a raw capture against the digest recorded in the manifest:

shasum -a 256 raw/lidar/20260902T1320_80858-0f4_columbus-broadway.pcapng

Consumers should prefer manifest fields over parsing semantic meaning back out of filenames. The filename is for navigation; the manifest is for joins.

Provenance and processing

The initial captures were segmented with pcap-split 0.5.1-pre31 and 0.5.1-pre32. Where the exact pipeline Git SHA is not available for a historical capture, pipeline_git_sha remains null rather than being reconstructed from circumstantial evidence.

The web scenes are derived views of the corresponding capture. They make the point clouds practical to inspect in a browser, but they do not replace the raw PCAPNG evidence. When the two serve different questions, use the raw capture as the record and the scene as the convenient view.

Capture-specific processing and source lineage belong in the manifest rather than in release-wide prose so that the dataset can grow without making older statements ambiguous.

Known limits

• Positions are approximate readings from a field map.
• Some historical capture metadata, including sensor_model and pipeline_git_sha, may be unknown and therefore null.
• Reference labels cover one capture, kirkham-9th (83 seconds). They are foreground-domain and partial: a return the labelling run's foreground left out is unlabelled. The corpus contains no ground-truth tracks, trajectories, or reference speeds.
• Scene, track, map, metric, report, and radar artefacts may be absent for a capture; kirkham-9th has no scene export.
• Filename timestamps are minute-resolution; the manifest preserves the full timestamp and offset.
• A capture may include an operator-approved short tripod nudge within the same site interval. Such overrides are recorded in export_provenance.
• Geographic coverage is opportunistic and expanding; the corpus should not be treated as a statistically representative sample of San Francisco streets.

These are limits, not invitations to quietly fill the gaps. Research data has enough uncertainty without adding confidence by typography.

Privacy and responsible use

velocity.report measures movement without cameras, faces, licence plates, or identity tracking. This dataset follows the same principle. It is intended for street-safety research, reproducible perception work, and community evidence, not for identifying or monitoring individuals.

Raw point clouds still describe activity in public space. Users should assess their own derived outputs before publication and avoid adding identity-bearing data from other sources. Privacy is not inherited automatically by every model that happens to read a privacy-preserving input.

Project

The dataset is produced by velocity.report, an open-source, local-first project for measuring traffic movement and building inspectable evidence for safer streets.

The processing code, data conventions, and research work are developed in the open. The dataset is expected to grow as additional stationary surveys are collected and as reproducible derived products become available.