Rangel

FindX

A web app that checks a photo against registered missing-person cases and, on a match, alerts the nearest police station.

Read the code
When
November 2023
Area
Four police stations in north-west Mumbai
Type
Django web app with face recognition
Stack
Python, Django, DeepFace, OpenCV, SQLite, Bootstrap
Problem
Someone who spots a missing person usually can't tell if they really are missing, or which police station to call.
Solution
Upload a photo. FindX compares the face with every registered case and routes a match to the right station automatically.
Result
A working citizen app and police portal: case registration, face matching, station alerts, messaging and PDF case reports.

The problem

Missing-person posters depend on someone recognising a face and then knowing what to do about it. A passer-by rarely knows whether the person in front of them is actually on a list, and even if they are sure, the case belongs to one station while the sighting is near another.

FindX splits the work into two sides. Citizens get one job: take a photo. Police register cases and receive the alerts.

How a scan works

  1. Upload a photo

    A citizen who thinks they've seen a missing person uploads a picture. The browser sends their location along with it.

  2. Find the face

    MTCNN, a face detector, finds and crops the face. If there is no clear face, the scan stops and says so instead of guessing.

  3. Compare it

    DeepFace turns the face into a VGG-Face embedding and measures its cosine distance to every registered missing person's photo. Those embeddings are computed once and cached, so a scan doesn't redo them.

  4. Alert the police

    If the closest face is near enough, FindX notifies two stations: the one nearest the sighting and the one that registered the case.

FindX upload page with an 'Upload the Image' file picker and a Submit button
Step 1: the upload page.
FindX result page reading 'Notification sent! Match found, cosine value 5.55e-16, nearest police station: Charkop'
Step 4: a match, its cosine distance, and the station that was alerted.

What counts as a match

Cosine distance runs from 0, the same face, upwards as faces differ. FindX only calls it a match below 0.3. Anything above is reported as no match. A false alarm would send police after the wrong person, which is why the cut-off is strict.

Cosine distance between two faces
0.3
0, same face1.0, very different

Match. Notify the nearest station and the case's station.

No match. Tell the citizen, notify no one.

Finding the station

The prototype covers four stations in north-west Mumbai: Gorai, Charkop, Kandivali and Borivali. FindX uses the haversine formula to measure the straight-line distance from the sighting to each one and picks the closest.

The four stations, at their real coordinates
GoraiCharkopKandivaliBorivaliSighting
  • Gorai 1.9 km
  • Charkop 0.7 km
  • Kandivali 2.5 km
  • Borivali 3.5 km

Station positions from the code; the sighting is an example.

The police portal

Each station signs in to its own account and gets:

Case registration
A detailed form: name, age, height, last known location, hair, skin tone, eye colour, clothing, physique, any dispute or suspect, a reward, a government ID document and the photo that gets matched against.
Case dashboard
Every case with its status, which the station moves between Active, Resolved and Closed.
Notifications
Matches sent to this station, either because the sighting was nearby or because it owns the case.
Station messages
A simple message thread between stations about a case.
PDF reports
A printable case report generated from the record with xhtml2pdf.

On the citizen side, the home page lists open cases, and Google Translate is built in so the app works in more than one language.