Litter Watch — Street Garbage Detection
Finds street garbage from a moving scooter and pins every pile to a real coordinate a municipality can act on.
A computer-vision system that finds garbage piles in street video shot from a moving scooter, pins each one to a real coordinate, and streams annotated frames to a live dashboard — with a geometry gate that filters out people, vehicles and street furniture.

- Industry
- Civic / Municipal
- Role
- ML & Full-Stack Engineer
- Type
- AI System
- Year
- 2025
01
Overview
Litter Watch detects garbage and trash piles in street footage filmed from a phone mounted on a moving scooter — the cheapest possible survey rig for an Indian city. A YOLO detector runs over sampled frames, and every confirmed pile is pinned to its own coordinate so the output is a map of where the problem actually is, not just a count.
The hard part is not detection but suppression. Street scenes are full of things that look like a trash pile to a naive detector: people crouching, parked vehicles, stones, tree litter, benches, posters and fences. A geometry gate filters these forbidden classes explicitly and reports how many boxes it rejected, so the confirmed count means something.
The same detector drives two faces — a FastAPI backend streaming an MJPEG feed to a Next.js dashboard with a live map and per-locality stats, and a CLI that batch-processes a video into an annotated file plus a JSON report.
02
Business challenge
Municipal garbage surveys are done on foot with clipboards — slow, expensive and stale by the time they are compiled. There was no cheap way to answer "where exactly is the garbage" across a city, and a naive detector run over street video reports every crouching person and parked vehicle as a pile.
03
Solution
A detection pipeline driven by phone video from a moving scooter: sampled frames run through a YOLO model trained on trash piles, a geometry gate rejects forbidden classes, and every surviving detection is pinned to a coordinate and streamed to a live dashboard.
What it does
04
Product screenshots
Screens from the application itself. Select any image to view it full size.
05
Architecture
How it fits together
- 01YOLOv8 detector with custom trash-pile weights, run through Ultralytics and OpenCV
- 02FastAPI backend exposing detection control, run statistics and an MJPEG frame stream
- 03Next.js dashboard with live feed, Leaflet coverage map, detection gallery and per-zone stats
- 04Geometry gate that filters forbidden classes and reports the rejection count for transparency
- 05Sandboxed video library — the detector resolves every job inside one folder and rejects escaping paths
- 06CLI batch path sharing the same detector, emitting an annotated video plus a JSON report
06
Engineering challenges
- Challenge
- The detector confidently boxed people, scooters, stones and posters as garbage — a false-positive rate that made the pile count meaningless.
- How it was solved
- Added an explicit geometry gate for forbidden classes and surfaced the rejected-box count in the UI, so suppression is visible rather than hidden. Green boxes show what was filtered; only red boxes are ever reported as garbage.
- Challenge
- Footage from a moving scooter means the same pile appears across many consecutive frames and would be counted repeatedly.
- How it was solved
- Sampled at a configurable frame interval and pinned confirmed detections to coordinates, so a pile resolves to one location rather than one count per frame it appears in.
07
Technology stack
- Frontend
- Next.jsReact
- Backend
- PythonFastAPI
- AI & Processing
- YOLOv8OpenCV
- Also used
- UltralyticsLeafletMJPEG streaming
08
Business impact
- 78%+
- Avg Confidence
- Live video
- Detection
- Geo-pinned piles
- Output
A survey that previously needed people walking with clipboards became a scooter ride: in a single run the system processed street footage live, confirmed piles at 78%+ average confidence, filtered the forbidden-class boxes, and dropped every confirmed pile onto a Ranchi coverage map with a per-locality breakdown.
09
Project status
What I took from it
- On a real-world detector, the filtering rules carry more product value than the model — and showing what was rejected is what makes the remaining count trustworthy.
- Constraining file jobs to a single sandboxed folder is worth doing on day one when the input is user-uploaded video.
Next step
Need a similar solution for your business?
Tell me what you are trying to build or fix. I will come back with how I would approach it, what it realistically takes and where the risks are.


