NightGuard: Low-Light Surveillance Detection System

CSE468 - Computer Vision Project | Group 8 Supervised by Dr. Mohammad Shifat-E-Rabbi, North South University

---

Overview

NightGuard is a real-time surveillance system designed to detect and identify objects in low-light CCTV footage. The system combines image enhancement techniques with deep learning-based detection to handle challenging nighttime conditions.

Pipeline Architecture

Raw CCTV Frame (Low-Light)
        │
        ▼
┌──────────────────────────────┐
│  Low-Light Enhancement       │  ← Anindya (Module 1)
│  (DL Ensemble / CLAHE)      │
└──────────────┬───────────────┘
               │ Enhanced Frame
               ▼
┌──────────────────────────────────────────────┐
│            Parallel Detection                │
│                                              │
│  ┌────────────┐ ┌────────────┐ ┌───────────┐│
│  │   Face     │ │   Human    │ │  Vehicle  ││
│  │ Detection  │ │ Detection  │ │ Detection ││
│  │ (Midhat)   │ │ (Abhishek) │ │ (Maisha)  ││
│  └────────────┘ └────────────┘ └───────────┘│
└──────────────────────────────────────────────┘
               │
               ▼
         Fused Results
   (Bounding Boxes, Labels, Confidence Scores)

View Project Presentation

---

Quick Start

# 1. Clone and install
git clone https://github.com/AbhishekKaisar/NightGuard-System.git
cd NightGuard-System
pip install -r requirements.txt

# 2. Download required weights (not included in repo due to size)
#    - support/onnx_weights/    → from Google Drive (see Weights section below)
#    - support/maisha_weights/  → included in repo

# 3. Run the full pipeline on a sample image
python3 main.py --input data/x1080.jpg --output results/output.jpg

Or open support/notebooks/NightGuard_Demo.ipynb in Google Colab for an interactive demo.

---

Team Members

Name	ID	Role	Module Folder
Anindya Saha Ani	2221105042	Low-Light Enhancement Lead	support/modules/enhancement/
Midhat Bin Shazzad	2222560642	Face Detection Lead	support/modules/face_detection/
Abhishek Kaisar Abhoy	2221140042	Human Detection Lead	support/modules/human_detection/
Maisha Tabassum	2222728042	Vehicle Detection Lead	support/modules/vehicle_detection/

---

Project Structure

NightGuard-System/
├── main.py                 # Main pipeline — run this to detect
├── README.md               # Project documentation
├── requirements.txt        # Python dependencies
├── data/                   # Sample test images (datasets)
├── support/                # All supporting code and models
│   ├── modules/
│   │   ├── enhancement/    # Anindya — Deep learning ensemble
│   │   ├── face_detection/ # Midhat — YOLOv8n-face detection
│   │   ├── human_detection/# Abhishek — YOLOv8n human detection
│   │   └── vehicle_detection/ # Maisha — YOLOv8n + RT-DETR
│   ├── notebooks/          # Jupyter notebooks
│   ├── maisha_weights/     # Fine-tuned vehicle detection weights
│   ├── onnx_weights/       # ONNX models for fast CPU inference
│   ├── export_onnx.py      # ONNX export script
│   └── tune_pipeline.py    # Hyperparameter tuning script
├── others/                 # Presentations, reports, demo video
├── results/                # Output images and evaluation metrics
└── docs/                   # Technical documentation

Modules

1. Low-Light Enhancement — Anindya Saha Ani

• Deep Learning Ensemble: Fuses four frozen base models (Zero-DCE, KinD, RetinexNet, Restormer Vision Transformer)
• Meta-Learner: U-Net Fusion Engine for dynamic spatial feature weighting
• Optimized Inference: ONNX Runtime FP16 on CPU, PyTorch on GPU — no GPU required
• Exposure Safety Check: Auto-fallback to CLAHE if the DL model over-exposes the image
• Downscaling: Images above 1080p are automatically downscaled before enhancement
• Deployment: Gradio web interface (app.py) for drag-and-drop inference

2. Face Detection — Midhat Bin Shazzad

• YOLOv8n with face detection weights
• CLAHE + Fast Non-Local Means Denoising preprocessing
• Dual-input smart selector (runs on both raw and enhanced, picks best confidence)
• Confidence improvement: ~0.42-0.65 (raw) → ~0.70-0.90 (enhanced)

3. Human Detection — Abhishek Kaisar Abhoy

• YOLOv8n pretrained model (person class)
• Gaussian blur preprocessing for noise suppression
• Sobel edge detection for structural analysis
• Interactive gamma correction slider for parameter tuning
• Confidence improvement: 0.41 (raw) → 0.77 (enhanced)

4. Vehicle Detection — Maisha Tabassum

• YOLOv8n baseline + fine-tuned on ExDark dataset (2,320 vehicle images)
• RT-DETR (transformer-based) fine-tuned for low-light vehicle detection
• Smart model selection: runs both pretrained and fine-tuned, picks best result
• Bounding box validation to filter bad detections
• Vehicle classes: Car, Bus, Bicycle, Motorcycle
• Best result: RT-DETR fine-tuned — 0.893 avg confidence

Dataset

This project uses the ExDark Dataset — a collection of low-light images across 12 object categories. The dataset is not included in this repository due to size constraints.

Download from Google Drive: ExDark Dataset

Original Source: ExDark GitHub

After downloading, place it in a Dataset/ folder at the project root:

Dataset/
└── ExDark_Dataset/
    └── People/        # 609 low-light images

Weights (Download Required)

Some weight files are too large for GitHub. Download from Google Drive and place in the correct folders:

Folder	Contents	Download
onnx_weights/	ONNX ensemble model for fast CPU inference	Google Drive
modules/enhancement/weights/	Pretrained base model weights + U-Net fusion	Google Drive

Note: maisha_weights/ (vehicle detection) and modules/face_detection/yolov8n-face.pt are already included in the repo. yolov8n.pt auto-downloads on first run.

Setup

# Clone the repository
git clone https://github.com/AbhishekKaisar/NightGuard-System.git
cd NightGuard-System

# Create virtual environment (optional but recommended)
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install all dependencies
pip install -r requirements.txt

# Download weights from Google Drive (see Weights section above)
# Place onnx_weights/ and modules/enhancement/weights/ in the project root

# Run the pipeline
python3 main.py --input samples/x1080.jpg --output results/output.jpg

Tech Stack

• Deep Learning: PyTorch, Ultralytics YOLOv8, RT-DETR, Restormer, RetinexNet, Zero-DCE, KinD
• Optimized Inference: ONNX Runtime (FP16 for CPU)
• Computer Vision: OpenCV
• Deployment: Gradio
• Data Processing: rawpy, NumPy, Pandas
• Languages: Python 3.8+
• Environment: CPU (ONNX) / GPU (PyTorch) / Google Colab

License

This project is developed for academic purposes as part of the CSE468 course at North South University.