Measure velocity, not identity
Street-level speed measurement for neighbourhood change-makers, researchers, and anyone who needs evidence instead of guesswork about what traffic does. Radar and LiDAR sensors measure speeds and movement: no cameras, no licence plates, no individual is identified, tracked, or recorded. The data belongs to the community that collected it.
- 📊 PDF reports: speed distributions, percentile statistics, before-and-after comparisons
- 🔒 Privacy by architecture: the hardware cannot collect what the design never asked for
- 📡 Radar speed measurement and LiDAR object tracking (working toward sensor fusion, combining both sensors)
- 🏠 Runs locally in your neighbourhood, offline-first
- 🔒 Open source and auditable, because trust should be verifiable
░░░░░░ ░░░░░░
░░░░░ ░░░░░░░░░░░ ░░░░░ :::. ░░░░░░░░░░ ░░░░░░░░░░░ ░░░░░ :::.
░▒▒▒▒▒░░░░▒▒▒▒▒▒▒▒▒▒▒░░░░▒▒▒▒▒ .:. ▒▒▒▒▒▒▒▒▒▒░░░░▒▒▒▒▒▒▒▒▒▒▒░░░░▒▒▒▒▒ .:.
░░▒▓▓▓▓▓▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒ .::.::: : ▓▓▓▓▓▓▓▓▓▓▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒ .::.:.: :
▒▒▓. . ▓▓▓▓████ ... . ▓▓▓▓ :....:::::. ▓▓▓. . ▓▓▓▓████ ... . ▓▓▓▓ :.::..::::
▓▓█ ..::████████ :.::: ████ :: : : . ::▓▓▓ ..::████████ :.::: ████ :: :: : ::
:.: : ::████:.:: : ::: ████ .: : . ... :.: ::::████:.:: : ::: ████ .::: .:::.
.:: . . ::.: . . . .:: . . ::.: :::: . .
. . .
▄▀▀▀▀▀▀▄
▄▄█▄▄▄ ▄▄▌
▌▐▀▀▀█ ▌
▀▀▀▀▀▀█████▀▀▀▀▀▀▀▀▀█▀▀█▀▐▀ ▄▄▀▀█▀▀▀▀▀▀▀█▀▀██▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀
▄▄▄▀▀▀░░ ░ ░ ░▒▓▄▄▀ ▄▄▀▀ ▄▓▀▀▀ ░▐▌ ▀▀▄▄▀▀▄▄▓▒ ▒ ░ ░ ░ ░ ░ ░
░ ░░ ░ ░ ▒ ▄▄▀▀▄▄▀▀▄▄▄▄▓░ ░▄ ░█ ▀▀▄▄▀▀▄▄ ▒ ░ ░ ░ ░ ░ ░
░ ░ ▒▄▄▀▀▄▄▀▀ ▀▄▀█░ ▄▄▀█▀░░ ░█ ▀▀▄▄▀▀▄▄ ▒ ░ ░ ░
░░ ▒▄▄▀▀▄▄▀▀ █▀▀▓▀▀▌ ▐▌░ ░░▒▌ ▀▀▄▄▀▀▄▄ ░ ░ ░ ░ ░
▄▄▀▀▄▄▀▀ ▓ ▓ ▌ █▀▀▀▀▀▀█ ▀▀▄▄▀▀▄▄ ▒░ ▒
▄▄▀▀ ▓▀▀▓▀▀▌ ▐▌░ ░ ▐▌ ▀▀▄▄▀▀▄▄ ░ ▒
▓__▓__▌ █░ █▄░ ▐▌ ▀▀▄▄▀▀▄▄ ░
▓ ▓ ▌ █░ ▐▌ █░ █ ▀▀▄▄▀▀▄▄
▄▄▄ ▄▄▄▄▄▄ ▄▄▄▄ ▓ ▓ ▌▒▒▒▒▒ ▒▒▒▒ ▄▄▄▄▄ ▄▄▄▄ ▄▄▄▄▄ ▄▄▄▄▄ ▄▄▄▄▄▄▄ ▄
█▀ ▄█████▀ ▄████▀ ▄████ ████ ████ ██████ ▀███▄ ▀████▄ ▀████▄ ▀██████▄
▄█████▀ ▄████▀ ▄████▀ ▄████ █████ ██████ ▀███▄ ▀████▄ █████▄ ▀██████
█████▀ ▄█████ ▄████▀ █████ █████ ██████ ████▄ █████ ▀█████▄ ▀████
████▀ ▄█████ ▄█████ ████▀ ▀████▄ ██████ ████▄ ▀████▄ ▀█████ ▀██
██▀ ▄█████▀ ▄█████ ▄████ █████ ██████ ▀████▄ ▀█████ ▀█████▄ ▀
▀ ▄█████▀ ▄█████▀ █████ ██████ ██████ ▀████▄ █████▄ ▀█████▄
░██ ░██ ░██ ░██
░██ ░██ ░██
░██ ░██ ░██ ░██████ ░██░████ ░████████ ░██░████████ ░████████ ░██
░██ ░████ ░██ ░██ ░███ ░██ ░██ ░██░██ ░██ ░██ ░██ ░██
░██░██ ░██░██ ░███████ ░██ ░██ ░██ ░██░██ ░██ ░██ ░██ ░██
░████ ░████ ░██ ░██ ░██ ░██ ░██ ░██░██ ░██ ░██ ░███
░███ ░███ ░█████░██ ░██ ░██ ░██ ░██░██ ░██ ░█████░██ ░██
░██
░███████
This is alpha software. It works. It is tested. It is not yet hardened. No authentication, no access control: an exposed instance will serve its data to anyone who asks and the PDF endpoint will fill up a hard drive if someone keeps clicking "Generate Report". Deploy on a private network or a dedicated VLAN with its own storage budget (or better yet, dedicated hardware like a Raspberry Pi). Do not point it at the open internet and hope for the best: hope is not a security architecture. See BACKLOG.md for current priorities.
Communities trying to make their streets safer face a familiar problem: everyone has an opinion about how fast cars go, but nobody has evidence. Council meetings run on anecdote. Speed surveys cost thousands and arrive months late. Meanwhile, someone's child is still crossing that road while motorists zip by.
velocity.report exists to close the gap between feeling unsafe and proving it.
A radar measures vehicle speeds. The logged data stays on a local device in someone's house. The reports are professional enough for a planning committee.
Evidence over opinion. Privacy over convenience. Community ownership over cloud dependency.
- Neighbourhood groups measuring speed on their street, with evidence instead of guesswork
- Community advocates building a case for traffic calming, with data strong enough for a formal submission
- Transport agencies and government departments running street speed surveys (whether for planning, compliance, or understanding what a road redesign actually did to driver behaviour) without the cost or lead time of traditional survey contractors
- Academics and researchers studying street-level vehicle behaviour with raw LiDAR point clouds, multi-object tracking, and replayable datasets
- Perception and robotics engineers exploring a transparent LiDAR pipeline: DBSCAN clustering, Kalman-filtered tracking, and rule-based classification; all tuneable and documented from raw UDP frames to classified tracks
- Before-and-after studies showing whether traffic calming interventions actually work
The system is built to measure traffic speed and protect vulnerable road users. That defines what we log, what we store, and what we report.
The software does not collect, store, or expose anything that identifies an individual or a specific vehicle. No licence plates. No faces. No make or model. Object classification tracks road user types: pedestrian, cyclist, car, bus. Not people. Not plates.
The data stays on a local device. Reports are generated locally. If personal data reaches a log, a response body, or an export, that is a bug, please report it.
See TENETS.md for the full set of non-negotiable principles.
Clarendon Avenue runs past an elementary school in San Francisco, designated a high injury road for years. Parents worried about vehicle speeds and whether the city's planned repaving would actually help.
When the city announced a quick-build to repave Clarendon, the Banshee team ran a baseline speed survey in June 2025 and sent the results to the city engineering team ahead of their October planning meeting. The city repaved in December 2025. A second survey in January 2026 showed what parents feared: speeds had climbed, not fallen. The repaving, funded to improve safety, had made the school run more dangerous.
The team presented the findings at a San Francisco City Hall street safety hearing in January 2026. The full PDF is available at banshee-data.com.
| Metric | Period 1 | Period 2 | Change |
|---|---|---|---|
| P50 speed | 30.54 mph | 33.02 mph | +8.1% |
| P85 speed | 36.94 mph | 38.70 mph | +4.8% |
| P98 speed | 43.05 mph | 44.21 mph | +2.7% |
| Max speed | 53.52 mph | 53.82 mph | +0.6% |
| Vehicle count | 3,460 | 2,455 |
Clarendon Avenue school zone survey: speed distributions, percentile statistics, and period-over-period comparison
| Component | What it does |
|---|---|
| Go server | Collects radar speed data and LiDAR point clouds, stores both in SQLite, serves the API. → cmd/, internal/ |
| macOS visualiser | Native 3D LiDAR point cloud viewer with object tracking, replay, and debug overlays. Apple Silicon. → tools/visualiser-macos/ |
| Web frontend | Data visualisation and interactive charts for recorded speed data. Svelte + TypeScript. → web/ |
| PDF generator |
Turns speed data into professional reports with charts, statistics, and proper formatting. Python + LaTeX, deprecated: PDF generation is moving into the Go server. → tools/pdf-generator/ |
Requires Go 1.25+, Node.js 18+, and pnpm. See CONTRIBUTING.md for full prerequisites.
git clone [email protected]:banshee-data/velocity.report.git
cd velocity.report
make build-web
make build-radar-localThe build produces velocity-report-local. Start it without a connected sensor:
./velocity-report-local --disable-radar --listen :8080The server creates a new SQLite database if one does not exist.
Open localhost:8080 to see the dashboard.
Use --db-path to point at an existing database elsewhere.
┌──────────────────┐ ┌──────────────────────────┐ ┌──────────────────┐
│ Sensors │────►│ velocity.report Server │◄───►│ SQLite Database │
│ (Radar / LiDAR) │ │ (Go) │ │ (sensor_data.db) │
└──────────────────┘ └──────────────────────────┘ └──────────────────┘
│ │
HTTPS :443 │ │ gRPC :50051
(nginx proxy) │ │
┌──────────────┴─┐ │
│ │ │
▼ ▼ ▼
┌──────────────┐ ┌───────────────┐ ┌─────────────────────┐
│ Web Frontend │ │ PDF Generator │ │ VelocityVisualiser │
│ (Svelte) │ │ ⚠ deprecated │ │ (macOS/Metal, gRPC) │
└──────────────┘ └───────────────┘ └─────────────────────┘
The web frontend and PDF generator connect over HTTP (:8080). The macOS visualiser uses gRPC (:50051) for streaming point cloud data. The Python PDF generator is deprecated: PDF generation is moving into the Go server. For the full architecture see ARCHITECTURE.md. Sensor fusion plans live in VISION.md.
Every contribution needs three commands:
make format # auto-fix all formatting (Go, Python, Web)
make lint # check formatting, code quality, and repo hygiene
make test # run all test suitesThese are the core quality checks for local development and CI, but pull requests may also run additional repository or documentation checks.
Check out docs/plans and BACKLOG.md for any work that catches your eye.
See CONTRIBUTING.md for dev environment setup, coding standards, and pull request workflow. All make targets are documented in COMMANDS.md.
The Go server runs as a systemd service on Raspberry Pi. See public_html/src/guides/setup.md for the complete setup guide.
The LiDAR pipeline already runs a full perception stack: DBSCAN spatial clustering, Kalman-filtered multi-object tracking with state and covariance estimation, and rule-based classification across eight object types (car, truck, bus, pedestrian, cyclist, motorcyclist, bird, and general dynamic). Radar provides independent Doppler-accurate speed. Today, both run in parallel.
The next stage fuses them into a single scene model: cross-sensor track handoff using Mahalanobis-distance gating, Bayesian evidence accumulation for persistent geometry, and canonical object refinement via streaming statistics. A lorry clipping a corner at 35 mph tells a different story to a bicycle at the same speed; radar captures the speed regardless, and fusion adds the geometry to tell the two apart.
The full plan is in docs/plans/lidar-l7-scene-plan.md and VISION.md.
LiDAR point cloud visualiser: real-time multi-object tracking with bounding boxes and velocity vectors
Native Metal renderer for live and recorded LiDAR data, with 96% bandwidth reduction through background caching and foreground-only streaming. Requires macOS 14+ and Apple Silicon.
make dev-macOpen the LiDAR Dashboard at localhost:8081 to replay captured point cloud data (.pcap files). See tools/visualiser-macos/README.md for controls and camera navigation.
- 📐 TENETS.md: the six tenets that guide the project: privacy, safety, evidence, local-first, simplicity, DRY
- 🔭 VISION.md: where this is going, and what it looks like when it gets there
- 🏗️ ARCHITECTURE.md: how the parts fit together: sensors, server, database, frontend
- ✨ MAGIC_NUMBERS.md: canonical constants for ports, limits, and tuning defaults
- 🔢 MATHS.md: the maths behind tracking, clustering, and classification, with no handwaving
- 🤝 CONTRIBUTING.md: how to contribute without producing an exciting new class of bug
- 🧱 COMMANDS.md: every make target, in one place
- 🎨 DESIGN.md: how the interface should look and why
- 🧭 DECISIONS.md: why things are the way they are
- 📊 COVERAGE.md: what is tested, what is not, and where the gaps live
- 🪵 CHANGELOG.md: what changed and when
- 📓 DEVLOG.md: engineering journal
- 📋 BACKLOG.md: prioritised work queue
- 🌲 MATRIX.md: test and validation surface coverage
- ❓ QUESTIONS.md: things we do not know yet
- 🛠️ DEBUGGING.md: when things go wrong, start here
- 🤝 CODE_OF_CONDUCT.md: how we treat each other
Join the Discord to discuss the project, get help, and help make streets safer.
Apache License 2.0. See LICENSE.