VisionSafe is an AI-powered video surveillance system designed to detect safe and unsafe activities in uploaded videos using computer vision.
The system analyzes videos, detects objects using YOLOv8, and highlights unsafe scenarios with bounding boxes.
It provides a web dashboard where users can upload videos and view processed results.
🎯 Key Features
🎥 Upload video through a modern web interface
🤖 AI-based object detection using YOLOv8
🟥 Detect unsafe objects (vehicles, suspicious objects, actions)
🟩 Detect safe activities (people)
📦 Bounding box visualization on detected objects
🎞 Processed video output generation
📊 Detection result displayed on dashboard
⚡ Fast inference using OpenCV and Ultralytics
🌐 Full-stack web application (React + Flask)
🧠 AI Detection Logic
The AI model classifies detected objects into Safe and Unsafe categories.
Detected Object Classification Person SAFE Car UNSAFE Motorcycle UNSAFE Bus UNSAFE Truck UNSAFE
Bounding boxes are drawn around objects with labels.
Example:
🟩 Green box → SAFE
🟥 Red box → UNSAFE
🏗 Project Architecture safe-usafe-activity-detection │ ├── client/ # React Frontend │ ├── src │ └── package.json │ ├── server/ # Flask Backend │ ├── routes │ ├── controllers │ ├── services │ └── app.py │ ├── Label/ # AI Model & Inference │ ├── inference.py │ ├── yolov8n.pt │ └── outputs │ ├── screenshots # UI screenshots │ ├── README.md └── .gitignore 🛠 Tech Stack Frontend
React
Vite
JavaScript
CSS
Backend
Python
Flask
Flask-CORS
AI / Computer Vision
YOLOv8 (Ultralytics)
OpenCV
NumPy
Development Tools
VS Code
Git
GitHub
Navigate to server folder:
cd server
Install dependencies:
pip install flask flask-cors ultralytics opencv-python numpy pandas mediapipe
Run backend server:
python app.py
Server will start at:
http://127.0.0.1:5000 3️⃣ Start Frontend
Open another terminal.
Navigate to client folder:
cd client
Install dependencies:
npm install
Run frontend:
npm run dev
Frontend will start at:
http://localhost:5173 📷 Project Screenshots Dashboard
Upload Video
Detection Output
🎞 Example Workflow
1️⃣ Upload a video through the dashboard 2️⃣ AI processes frames using YOLOv8 3️⃣ Objects are detected 4️⃣ Bounding boxes are drawn 5️⃣ Safe / Unsafe result is displayed 6️⃣ Processed video is returned to user
🚀 Future Improvements
Real-time CCTV monitoring
Activity recognition (yoga, smoking, fighting)
Multi-person tracking
Suspicious object detection
Alert system
Cloud deployment
Mobile app integration
📊 Model Evaluation Results
The activity detection system was evaluated on a labeled dataset of 2,551 test samples using a Random Forest classifier trained on extracted pose features.
Overall Performance Metric Score Accuracy 97.14% Precision 97.03% Recall 97.14% Weighted F1 Score 97.06% Macro F1 Score 48.00%
The high weighted metrics indicate strong performance across the dominant classes, while the macro score highlights the impact of class imbalance.
📈 Per-Class Performance Class Precision Recall F1 Score Support Class 0 0.93 0.63 0.75 105 Class 1 0.97 1.00 0.99 2414 Class 2 1.00 0.10 0.18 31 Class 3 0.00 0.00 0.00 1
Observation
Class 1 dominates the dataset
Rare classes reduce macro-average score
Model performs strongly on frequent activity types
🔍 Confusion Matrix Predicted C0 C1 C2 C3 Actual C0 66 39 0 0 Actual C1 5 2409 0 0 Actual C2 0 28 3 0 Actual C3 0 1 0 0
Interpretation:
Most predictions fall correctly on the diagonal
Misclassifications occur mainly between C0 and C1
Extremely rare classes are harder for the model to learn
🧠 AI Processing Pipeline
The VisionSafe detection pipeline consists of the following stages:
Video Input │ ▼ Frame Extraction (OpenCV) │ ▼ Object Detection (YOLOv8) │ ▼ Pose Feature Extraction │ ▼ Activity Classification │ ▼ Safe / Unsafe Decision │ ▼ Bounding Box Rendering │ ▼ Processed Video Output ⚙️ Model Architecture Object Detection
Model used:
YOLOv8n
Advantages:
Lightweight
Fast inference
Good real-time performance
Activity Classification
Model:
Random Forest
Input features:
Pose landmark angles
Motion patterns
Bounding box context
🧪 Inference Performance Metric Value Average inference time ~30-40 ms per frame Model size ~6 MB FPS (CPU) ~20-25 FPS Detection framework Ultralytics YOLOv8 📊 System Performance Summary Component Technology Object Detection YOLOv8 Pose Analysis MediaPipe Activity Classification Random Forest Backend Flask Frontend React + Vite 🧩 System Architecture Diagram React Dashboard │ ▼ Flask Backend API │ ▼ AI Inference Engine ┌───────────────┐ │ YOLOv8 Model │ │ Pose Analysis │ │ RF Classifier │ └───────────────┘ │ ▼ Processed Video Output 🧠 Why This Project Is Important
VisionSafe demonstrates how AI can assist public safety monitoring by automatically identifying risky activities in video feeds.
Potential applications:
Smart city surveillance
Public transport monitoring
Campus safety systems
Traffic monitoring
👩💻 Author
Kareena Kumari
GitHub Profile https://github.com/kareena0229
⭐ Support
If you found this project useful, please consider giving it a star ⭐ on GitHub.
It helps others discover the project.