RAIDEN: Real-world Attack and Incident Detection ENgine

A simple browser-based interface to run YOLO and RF-DETR on images and videos. It features real-time multi-class tracking with persistent IDs, vehicle–person collision detection (time-to-collision based), and gun/explosion threat heatmaps.

This project was developed as part of the Master's thesis in Computer Science at the University of Salerno, carried out during a research internship at the Jheronimus Academy of Data Science (JADS) in 's‑Hertogenbosch, The Netherlands 🎓.

🚀 Features

• Image and video inference (JPG/PNG and common video formats)
• YOLO and RF-DETR model support
• Real-time per-class tracking with stable IDs and on-frame overlays
• Vehicle–person collision detection using geometric TTC heuristics
• Threat heatmap overlay for guns and explosions
• Gradio-based UI, runs locally in your browser

📁 Project Structure

MasterThesis/
├── backend/
│   ├── inference/
│   │   ├── yolo_inference.py          # YOLO inference pipeline
│   │   └── rfdetr_inference.py        # RF-DETR inference pipeline
│   ├── temp/
│   │   └── outputs/
│   │       ├── yolo/
│   │       │   └── video_results/     # YOLO processed videos
│   │       └── rfdetr/
│   │           └── video_results/     # RF-DETR processed videos
│   ├── stats/                         # JSONL logs (e.g., collisions.jsonl)
│   ├── weights/
│   │   ├── yolo/weights_yolo.pt       # YOLO weights (required)
│   │   └── rf_detr/weights_rfdetr.pt  # RF-DETR weights (required)
│   └── main_gradio.py                 # Main UI
├── requirements.txt                   # Python dependencies
├── run_server.py                      # Launcher
└── README.md                          # This file

🛠️ Installation

Prerequisites

• Python 3.8+
• pip

Setup

1. Clone the repository

git clone <repository-url>
cd MasterThesis

2. Install dependencies

pip install -r requirements.txt

3. Place model weights

• YOLO: put weights at backend/weights/yolo/weights_yolo.pt
• RF-DETR: put weights at backend/weights/rf_detr/weights_rfdetr.pt

4. Run the server

python run_server.py

5. Open the browser

• URL: http://localhost:7860

🎯 Usage

1. Upload a video file (or supported media)
2. Choose the model (YOLO or RF-DETR) and set confidence threshold
3. Start inference; the processed video and status message will appear in the UI

Supported Formats

JPG, JPEG, PNG, BMP, TIFF, MP4, AVI, MOV, MKV, WMV, FLV, WebM

Outputs

• Processed videos are saved under:
  • backend/temp/outputs/yolo/video_results/ (YOLO)
  • backend/temp/outputs/rfdetr/video_results/ (RF-DETR)

⚙️ Configuration

• Confidence threshold: controlled via the UI slider
• Advanced parameters (e.g., collision thresholds, tracker settings, threat weights) are defined in:
  • backend/main_gradio.py wrappers (default values passed to inference)
  • backend/inference/yolo_inference.py and backend/inference/rfdetr_inference.py

🔧 Development

• backend/main_gradio.py — Gradio interface and app wiring
• backend/inference/yolo_inference.py — YOLO inference pipeline, tracking, overlays, logging
• backend/inference/rfdetr_inference.py — RF-DETR inference pipeline, tracking, overlays, logging
• run_server.py — launcher and temp directory management

Logs and debug

• Console logs are printed during runtime
• Check temporary files under backend/temp/ and outputs under backend/temp/outputs/