Unified Safety AI Camera demo
IUB CAPSTONE PROJECT / 2026PROJECT RECORD

UnifiedSafety AI.

See risk sooner.
Respond with confidence.

Use existing CCTV for face detection and recognition, plus alerts for weapons, fights, falls, fire and smoke.

Explore the models Open camera demo
Record
IUB capstone project · 2026
Modules
Face · weapon · fire · fall · fight
Preview
Six camera feeds

Multi-camera preview

06 FEEDS · DEMO MEDIA
CAM 01

Person detection

CAM 02

Face recognition

CAM 03

Weapon detection

CAM 04

Fall detection

CAM 05

Fire and smoke

CAM 06

Aggressive activity

BUILT ATThe Islamia University of BahawalpurFaculty of ComputingDepartment of Artificial Intelligence

DETECTION STUDIES

What the system sees.

Scroll through the five modules, from face recognition to motion between people.

01

Detection study

Face detection
& recognition

Who is this?

  1. Input
  2. Detect
  3. Face mesh
  4. Align
  5. Embedding
  6. Match
  7. Result
Recorded walkthrough of the recognition pipeline · 25 s · both faces are members of the team

Detects faces and compares them with enrolled profiles, showing recognized identities and flagging unknown faces for review.

  • Face matching
  • Unknown-person warning

99.83%LFW benchmark accuracy

Model
InsightFace buffalo_l
Dataset
WebFace600K
IJB-C
97.25%
South Asian faces
93.16%

The recorded walkthrough takes two team members' faces through detection, 106 landmarks, alignment, a 512-number embedding and a comparison against every enrolled record.

  • 106landmarks per face
  • 512numbers per embedding
  • 93.3%match on face 1 in this demo
  • 93.2%match on face 2 in this demo
02

Detection study

Weapon
detection

Is anyone armed?

System output, untouched · left: camera with overlays · right: the person card the system builds as it goes

Identifies pistols and knives, associates them with a person, and confirms detections across multiple frames.

  • Pistols + knives
  • Multi-frame confirmation

90.5%Weapon precision

Model
YOLOv8m
Dataset
3 merged weapon datasets
Recall
86.0%
F1
88.2%
mAP@50
89.9% (pistol 94.3%, knife 98.0%)
mAP@50–95
80.5%
events.csv — written by the system during this clip
00:00.7IDENTIFIEDstudent mubeen recognised by face
00:02.2WEAPONmubeen is holding a PISTOL (conf 0.93)

Class 0 — HK USP pistolClass 1 — M9 bayonet

In this clip the system recognises Mubeen by his face first. A second and a half later, it knows he is armed. The two 3D models are the object classes the weapon module looks for.

03

Detection study

Fire
detection

Is something burning?

Staged fire test · recorded footage

Looks for flame and smoke cues, then assigns a clear risk level before alerting.

  • Fire + smoke
  • 1.5 sec confirmation

78.0%Fire precision

Model
YOLO11s
Dataset
Kaggle fire/smoke dataset
Recall
69.7%
F1
73.6%
mAP@50
77.9%
mAP@50–95
44.9%
04

Detection study

Fall
detection

Did someone go down?

Staged fall test · recorded system output

Tracks each person independently and escalates an alert when someone remains down.

  • Moderate · 4 sec
  • Severe · 15 sec

83.3%Fall precision

Model
SAFER CNN1D
Dataset
SAFER-Activities
Recall
85.0%
F1
84.1%
Accuracy
91.2% on unseen camera views

The overlay marks the detected fall and shows running counts of people and falls.

05

Detection study

Fight
detection

Is this turning into a fight?

System output · staged by the team · 6 s

Reads movement between nearby people over a short sequence and marks those involved for review.

  • Movement analysis
  • Person tracking

93%Fight module · validation accuracy

Model
5 BiLSTMs combined
Dataset
SCFD, RLVS, RWF-2000, BusViolence + crowd footage
fight_events.csv — written by the system during this clip
00:04.4FIGHTbetween track 1 and track 2 (score 0.84)

Watch the strip at the bottom of the frame: the fight score climbs from 0.00 while the two team members square up, then crosses the line.

THE PROJECT

Security teams can’t watch
every screen.
AI can help.

CCTV is everywhere. Attention isn’t.

Unified Safety AI checks camera feeds for urgent events, then sends a snapshot to a live dashboard so staff can assess what happened and decide how to respond.

It is built to work with existing cameras and process video on a local computer.

Follow the signal
OVERHEAD CROSSING FEEDRECORDED CLIP
1LIVE ALERT DASHBOARD
0NEW CAMERAS REQUIRED
DESIGNED FORFaster human responseAI helps surface an event. People make the call.

FROM CAMERA TO ACTION

Make the signal
actionable.

A short path from camera feed to a human-reviewed alert.

Built for local processingCore detection can run on site without sending continuous video to the cloud.
CONCOURSE FEEDRECORDED CLIP

  1. Camera feed

    Connect a webcam, IP camera, or video file.

  2. AI checks frames

    Specialist models look for a safety event.

  3. Confirm the event

    Temporal checks help filter one-frame noise.

  4. Alert for review

    A snapshot and event details reach the dashboard.

MEASURED · FRAME BUDGET

62 ms per frame

The system's own timing report from the fight clip: 230 frames in 14 seconds, 15.9 frames a second, with every model checking every frame on one local computer.

  • Find people34 ms
  • Weapons9 ms
  • Fire & smoke6 ms
  • Faces4 ms
  • Fights2 ms
  • Read frame2 ms
  • Draw & save5 ms

DEPLOYMENT & POTENTIAL USE

Security for places
people gather.

Our models have been tested in real-time scenarios. Face recognition is working in our university department. Weapon and fire detection are working with a client abroad. The settings below are potential environments for supervised pilots.

Passenger aircraft at a gate, seen from the terminal01

Airports & terminals

Concourses, entrances and restricted-area approaches

Passenger train alongside a railway platform02

Transit networks

Rail stations, bus terminals and transport hubs

Red brick academic building on a tree-lined campus03

Campuses

Universities, schools, hostels and research facilities

Emergency department entrance at a hospital04

Healthcare & care

Hospitals, clinics and elder care facilities

Spectator stands surrounding a stadium pitch05

Public venues

Tickets, event halls, convention and exhibition centres

Shoppers inside a multilevel retail centre06

Commerce

Banks, markets, malls and retail centres

Machinery and work areas inside an industrial facility07

Industry

Factories, warehouses, utilities and control rooms

Stone facade and tower of a municipal building08

Civic facilities

Government buildings and other public service sites

PRIVACY & HUMAN REVIEW

Model alerts should support trained staff, who review events before taking action. Performance varies with camera resolution, viewing angle, and lighting. Use appropriate privacy safeguards.

DEVELOPERS BEHIND THE PROJECT

Built with care.
Still growing.

The principal investigator and co-investigators behind Unified Safety AI at IUB.

Principal Investigator (PI)

Dr. Ghulam Gilanie

Assistant Professor

Department of Artificial Intelligence, IUB

Co-PI

M. Mubeen Mazhar

BS Artificial Intelligence · 7th semester

Computer vision · Deep learning · AI model integration

Google AI Professional Certificate

Co-PI

Muhammad Ahmed

BS Artificial Intelligence · 7th semester

Computer vision, deep learning & NLP

IBM AI Professional Certificate

A WIDER AI PORTFOLIO

Curious by
nature.

We have explored projects across computer vision, vision-language systems, recommendation, and deep learning.

  • APothole detection
  • BLicense plate recognition
  • CHeart disease prediction
  • DAI voice assistant
  • EText or 2D image to 3D generation
  • FAutonomous driving in CARLA

GOOD QUESTIONS

Clear answers
matter.

Common questions about the prototype.

Do we need new cameras?

No. The prototype can work with an existing IP camera, webcam, or video file.

Does it need an internet connection?

Detection runs on the local computer, so an internet connection is not required for the core workflow.

Is all video recorded?

No. The project is designed not to save continuous video. Alert photos and records of flagged events may be kept.

Is it 100% accurate?

No system is 100% accurate. We have tested our models in real-time scenarios and seen strong accuracy. Face recognition is working in our university department, while weapon and fire detection are working with a client abroad. Results still depend on camera resolution, viewing angle, and lighting, so trained staff should review every alert.

UNIFIED SAFETY AI / IUB

See what a camera
can notice.

Meet the team and learn how Unified Safety AI is being built at IUB.

Open camera demo Meet the team

Official project contact details coming soon