Violence Detection & Suspect Tracking System

This repository contains a multi-camera violence detection and suspect tracking system. It leverages YOLO for object detection, OSNet for Person Re-Identification (ReID), and a CNN-GRU model for violence detection. The system is designed to simulate a control room environment where events are detected, and suspects are tracked across multiple camera feeds.

🚀 Features

• Multi-Camera Simulation: Simulates a network of cameras with overlapping fields of view.
• Real-time Violence Detection: Uses a CNN-GRU model to detect violence in video streams.
• Suspect Tracking (ReID): Tracks suspects across different cameras using OSNet-based feature extraction.
• Event Orchestration: Manages active events, camera handoffs, and suspect locking logic.
• Visual Control Room: A GUI to visualize camera feeds, alerts, and tracking status.

📂 Project Structure

.
├── visual_simulation.py        # Main entry point for the visual simulation
├── train_hockey_gru.py         # Script to train the violence detection model (Hockey Dataset)
├── evaluate_ucf_crime.py       # Script to evaluate the model on the UCF-Crime dataset
├── test_yolo_pipeline.py       # Test script for the YOLO pipeline
├── requirements.txt            # Python dependencies
├── orchestration/              # Core logic for orchestration and dispatching
│   ├── engine.py               # Orchestration Engine (Event & Camera Management)
│   ├── reid/                   # ReID module
│   │   ├── feature_extractor.py # Feature extraction using OSNet
│   │   └── osnet.py            # OSNet model definition
│   └── ...
├── violence_detection_model.pth # Pre-trained Violence Detection Model
├── yolo26n.pt                  # YOLO model weights (Custom or v8)
└── ...

🛠️ Installation

1. Clone the repository:

   git clone <repository_url>
   cd <repository_name>

2. Create a virtual environment (optional but recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

4. Verify Model Weights: Ensure the following model weights are present in the root directory:

   • yolo26n.pt (or yolov8n.pt depending on configuration)
   • violence_detection_model.pth

🖥️ Usage

1. Run the Visual Simulation

This is the main demo of the system. It launches a window showing multiple camera feeds and simulates violence detection and suspect tracking.

python visual_simulation.py

• Controls:
  • Press q to exit the simulation.
  • Click on a bounding box to provide feedback (mark as false positive).

2. Training the Violence Detection Model

To train the CNN-GRU model on the Hockey Fight Dataset:

python train_hockey_gru.py --dataset hockey --data_root /path/to/HockeyDataset --epochs 20

3. Evaluation

To evaluate the trained model on the UCF-Crime dataset:

python evaluate_ucf_crime.py

Note: You will need the UCF-Crime dataset structure setup locally.

4. Testing YOLO Pipeline

To test the object detection pipeline on a single video:

python test_yolo_pipeline.py path/to/video.mp4

🧩 System Architecture

• Orchestration Engine: Manages the state of the entire camera network, handling event triggers and spatial handoffs.
• Dispatcher: Processes frames using a combination of YOLO (detection), OSNet (ReID features), and GRU (activity recognition).
• Feature Extractor: Extracts L2-normalized feature vectors for person re-identification.

📝 Configuration

• Camera Network: defined in orchestration/cameras_network.json.
• Model Parameters: logic for thresholds and model paths can be found in orchestration/engine.py and respective training scripts.