🚨 Real-Time Panic Detection System using Computer Vision

A real-time panic and abnormal crowd behavior detection system built using Computer Vision, Optical Flow, and YOLO.
The system detects sudden panic situations by analyzing motion intensity, direction changes, and velocity anomalies from CCTV or video feeds.

---

Problem Statement In crowded public spaces (stations, malls, religious events), panic situations escalate within seconds and often go unnoticed until damage occurs.
This project aims to automatically detect panic-like behavior early using live camera feeds, enabling faster intervention.

Solution Overview This system combines:

• Human detection using YOLO
• Motion analysis using Optical Flow
• Behavioral heuristics to detect panic

If there is:

• Sudden spike in motion velocity
• Abrupt direction changes
• Abnormal crowd flow

The system raises a panic alert.

Tech Stack

• Python
• OpenCV
• YOLO (Ultralytics)
• Optical Flow (Farneback)
• NumPy
• Real-time video processing

System Architecture

1. Video Input
   • CCTV feed / recorded video / webcam
2. Human Detection
   • YOLO detects people in each frame
3. Motion Analysis
   • Optical Flow calculates movement vectors
4. Panic Detection Logic
   • Sudden velocity or direction change triggers panic
5. Alert Syste
   • Visual / audio alerts (configurable)

Features

• Real-time processing
• Works on CCTV and live webcam
• Lightweight and fast
• Modular panic logic (easy to tune)
• No training required (rule-based + CV)

Installation

git clone https://github.com/jinay-k-jain/Panic_detector_CCTV.git
cd Panic_detector_CCTV/test_3
pip install -r requirements.txt
python real.py

Run on Video File python real.py --video path/to/video.mp4

Project Structure ├── real.py
├── alarm.py
├── requirements.txt
├── assets/
├── README.md

Panic Detection Logic (High Level) Track optical flow between frames Measure: Average motion magnitude Direction variance If values cross a threshold → panic detected Thresholds are configurable inside the code.

Limitations: Rule-based (not deep behavioral learning) Extreme camera shake may cause false positives Works best in medium-to-high crowd density

Future Improvements ML-based behavior classification Crowd density heatmaps Integration with control rooms Automatic alert dispatch to authorities Edge deployment (Jetson / CCTV NVRs)

Use Case Public events Railway stations Religious gatherings Stadiums Malls & metros

Author: Jinay Jain B.Tech, IIT (ISM) Dhanbad Smart India Hackathon Winner