Vadodara Access Map

A free, open-source map of physical-mobility barriers (broken or missing ramps, blocked footpaths, missing kerb cuts, no tactile paving, steps with no alternative) in Vadodara and its outskirts, India. It is a remote audit (street-level photos and OpenStreetMap data) plus a small set of photos taken by Om Dorkar. It is not a full survey: see docs/METHOD.md for what was and was not done.

Why Vadodara

Local news has reported for years that the city is hard to get around for people with disabilities:

• "Vadodara's buildings are not disabled-friendly", Times of India, 24 July 2014. Replies to RTI requests showed that most buildings in the district lacked ramps, accessible lifts, parking, toilets and signage.
• "City yet to adapt to needs of the disabled", Times of India, 4 July 2014. Apart from the collector's office, public buildings and places remained difficult to access, despite repeated petitions from NGOs and disability groups to the Vadodara Municipal Corporation.
• Motorised stair-climber wheelchairs launched at Vadodara railway station, DeshGujarat, 3 November 2022 (also covered by the Times of India). This is a welcome step, but the report notes it was needed because lifts and escalators could not be provided on the other platforms.

Those reports are now years old. This project asks a simple question: what is it like today, street by street? It answers with photos and locations anyone can check.

Live map: https://omdorkar.github.io/vadodara-access-map/ · Built by Om Dorkar · OpenStreetMap: Om Dorkar

How the AI works

This map pairs people with machine vision, and the machine learns from the people. Full detail in docs/LEARNING_LOOP.md.

1. Photos, no camera needed. Street-level photos from Mapillary, Panoramax and geotagged Wikimedia Commons photos (all openly licensed), plus Om's own photos. tools/learn/fetch_images.py.
2. Footpath rule. Mapillary labels every photo (road, sidewalk, kerb, crossing...). tools/scan_core.js asks: is there a road, and is a footpath visible? Whole stretches of street are coloured green or red (data/ai_streets.js); a road with no photo stays uncoloured.
3. Barrier detector. Open models run locally: CLIP embeds each photo, SegFormer measures road and sidewalk pixels, and a small classifier trained on Om's labelled photos, confirmed votes and the footpath rule proposes footpath present? and barrier type. Every photo is also matched to the most similar confirmed barrier photo. tools/learn/score_images.py → data/ai_barriers.js.
4. People check the AI. label.html shows one photo at a time with the AI's proposal and the look-alike confirmed photo. Y / N / can't tell. Two agreeing votes make a confirmed label.
5. It learns and measures itself. Every Monday (.github/workflows/learn.yml) the detector refits on the confirmed labels, re-scores every photo, and rewrites METRICS.md with cross-validated accuracy. The footpath scan (ai-scan.yml) runs the same morning.

Current state (see METRICS.md): 553 open photos scored; 0 confirmed votes yet; detector trained on 11 field photos + 331 weak footpath-rule labels; footpath present? at ~75% cross-validated agreement with the rule; barrier types not yet callable, so the map proposes look-alikes of Om's confirmed photos instead.

Footpath scan, first run (25 Oct 2026): 721 photos; 288 judged as built-up streets in daylight; a footpath was detected in 142 (49%) and not in 146. 13 zebra crossings were found. By area: Vadodara Bypass 32% (132 photos), Manjalpur 81% (90), Wadi 44% (32), Nava Bazaar 44% (27); other areas had too few photos. In a spot check of 12 "no footpath" dots, 8 clearly had none and 4 were unclear. The model almost never detects kerb ramps, so crossings are flagged to check, not counted as barriers. Coverage is mostly main roads photographed from vehicles.

Use it for another city

Everything city-specific is in data/city.js: the name, the box to scan, where the map opens and the time zone. Fork the repository, change that file, run the Action once (Actions tab, "AI street scan", Run workflow) and turn on GitHub Pages. The text on the page still mentions Vadodara in places and would need editing.

What's on the map

• AI street scan (small dots, on by default): green = footpath seen, red = no footpath seen on a built-up street, grey = open road, purple ring = zebra crossing to check for a ramp.
• Field survey pins (square): Om's own photos, each credited. A few show the setting but not the barrier itself, and their notes say so. Tap one for the photo, severity, details and location.
• Your reports (ringed): reports a visitor makes. They are sent for review and a copy stays in their browser (exportable as JSON).
• Documentation only (grey, dotted): photos that show no footpath barrier (Canal Road) or were taken far away (Pavagadh Hill). Not counted.
• AI-suggested (unverified) (purple diamonds, off by default): open street photos the detector proposes, or that look like one of Om's confirmed barrier photos. Nobody has checked them on the ground. They are not findings and are not counted. Check them and the detector learns.
• Mapillary photos (green dots): real street-level imagery (CC-BY-SA). You can mark a barrier you see in a photo.
• Already known in OSM (green / orange / red dots): places and kerbs where OpenStreetMap mappers recorded wheelchair access, kerb height or tactile paving. These are other people's data, shown as context.
• OSM possible gaps (purple dashes): mapped footways and crossings missing kerb / wheelchair / tactile_paving tags. This shows missing data, not a confirmed barrier, with a link to answer it in MapComplete. Off by default; switch it on under Layers & filters.
• Community reports (double ring): visitor reports shared to the database and approved by the project (see docs/SHARED_REPORTS.md).
• Street report: a printable, street-by-street list of the surveyed barriers with photos, locations and suggested fixes, for sending to the city.
• Street / satellite background (OSM and Esri World Imagery), plus filters by type and severity.

How to report a barrier

1. Tap Report a barrier and pick the closest type (missing ramp, broken ramp, blocked footpath, damaged footpath, no tactile paving, steps only, poor lighting).
2. Move the map so the pin sits exactly on the barrier, or tap Use my location.
3. Answer the short questions: severity, one detail question, an optional note and photo.

Where OpenStreetMap has a standard tag, the answer is stored as that tag (kerb=raised|lowered|flush, tactile_paving=no|incorrect, wheelchair=no|limited, lit=no). Verified points can then be added to OpenStreetMap, for example with MapComplete OnWheels.

Run it

Open index.html in a browser. There's no build step and nothing to install. If the street map tiles show "access blocked" when opened as a file, serve the folder instead, e.g. python3 -m http.server and visit http://localhost:8000. (OSM tile servers want a referring website.)

For the Mapillary layer, paste a free client token from https://www.mapillary.com/dashboard/developers into MAPILLARY_TOKEN in index.html, or into the box under Layers & filters.

Add survey photos

See PHOTO_GUIDE.md. In short: put photos in photos/originals/, notes in photos/notes.csv, then run

osascript -l JavaScript tools/process_photos.js "$PWD"

(The photo credit defaults to Om Dorkar; pass a different name as a second argument if someone else took the photos.) (macOS only; uses built-in ImageIO, no installs). Originals are git-ignored. Only metadata-stripped web copies are published.

Files

Path	Purpose
index.html	The map (Leaflet 1.9.4 from cdnjs). Settings for Mapillary, OSM username and Supabase are at the top of its script
report.html	Printable street report, generated from the survey data
label.html	Check-the-AI page: one photo, the AI's proposal, Y/N votes that become training labels
data/survey.js	Field survey records (generated, hand-editable: type, severity, note, tags, street)
data/vocab.js	Barrier types, severity levels, questions and suggested fixes, shared by both pages
data/city.js	City name, scan box, map start and time zone (the only file to change for another city)
data/ai_scan.js, data/ai_streets.js	Footpath scan results per photo and per street (generated)
photos/web/	Published survey photos and thumbnails
tools/process_photos.js	Reads photo GPS, makes web copies, looks up street names, writes data/survey.js
tools/scan_core.js, tools/ai_scan*.js	Footpath scan from Mapillary detections
tools/learn/	Learning loop: fetch open photos, score, learn from votes, export dataset (docs/LEARNING_LOOP.md)
data/ai_barriers.js	Detector proposals, one row per photo (generated)
labels/gold.csv	Confirmed labels from votes (generated); labels/checks_export/ for exported vote files
dataset/	Open dataset: barriers CSV/GeoJSON, AI proposals, MapRoulette challenge (generated)
METRICS.md	Detector accuracy, rewritten every run
supabase/schema.sql	Database tables and access rules for shared reports and AI checks
docs/SHARED_REPORTS.md	Setting up and reviewing shared reports
docs/LEARNING_LOOP.md	How the detector learns from votes, and its limits

Keeping it running

• Automatic: every Monday GitHub Actions re-runs the footpath scan and the learning loop and commits new results (author "ai-learn bot"). Both can be run by hand from the Actions tab.
• Repository secrets (Settings → Secrets and variables → Actions): SUPABASE_URL and SUPABASE_KEY (the public anon key) let the learning loop read votes; MAPILLARY_TOKEN is optional.
• Weekly, by hand: approve or reject pending visitor reports in Supabase (docs/SHARED_REPORTS.md).
• If Supabase pauses the free project after a quiet week, sign in and click Restore project. Until then the map still works and the learning loop falls back to exported votes.
• Everything used is free for a public repository.

Limitations

Visitor reports are reviewed for plausibility, not verified on the ground. OSM coverage of footways here is thin. Satellite imagery can't show ramp condition. Phone GPS is accurate to about 5–15 m.

Roadmap

1. Field survey: 10–15 barriers with geotagged photos, plus a Mapillary street-level sequence of the route (in progress).
2. Contribute verified kerb, tactile-paving and entrance data to OpenStreetMap, from the ground survey only.
3. Done (31 Oct 2026): shared, moderated storage (Supabase) for condition problems OSM doesn't map, such as broken ramps and blocked footpaths. Visitor reports go to review and appear for everyone once approved; votes from label.html are shared.
4. Talk to wheelchair users and a local disability organisation; send a street-by-street report to the Vadodara Municipal Corporation.
5. Built: a small ramp/obstacle detector trained only on suitably licensed photos, as a learning loop (docs/LEARNING_LOOP.md): it proposes, people confirm, it refits weekly. Next: fine-tune a segmentation model once confirmed labels pass ~500.

Prior work

The method follows the University of Alberta's Atlas / Click&Push project, which combines crowdsourced barrier reports with camera and GPS data. The reporting questions follow MapComplete OnWheels.

Licence

Code: MIT (LICENSE). Survey data and photos: CC-BY-SA 4.0 (LICENSE-DATA.md). Third-party data keeps its own licence (DATA_SOURCES.md).