🛰️ Portable Indigenous Multi-Modal Counter-UAS System (Vardan)

AI-Powered Lightweight Multi-Sensor Drone Detection Framework for Edge Devices

---

🎓 Final Year B.Tech Research Project

Department of Computer Science & Engineering
Amrita Vishwa Vidyapeetham

"Towards an Indigenous, Portable, and Energy-Efficient Counter-UAS Detection System using Multi-Modal Artificial Intelligence."

---

📖 Project Overview

The rapid increase in the use of Unmanned Aerial Vehicles (UAVs) has introduced security threats through unauthorized drone operations. This project develops Vardan, a lightweight, portable, multi-modal counter-UAS framework. By fusing inputs from Radio Frequency (RF), Acoustic waveforms, and Computer Vision feeds, Vardan increases detection robustness under noise while remaining computationally optimized for microcontroller and embedded-device deployment.

System Architecture Diagram

                    Portable Counter-UAS System
                  ┌───────────────────────────────┐
                  │      Incoming Drone           │
                  └──────────────┬────────────────┘
                                 │
         ┌───────────────────────┼────────────────────────┐
         │                       │                        │
         ▼                       ▼                        ▼
    RF Detection            Acoustic Detection      Vision Detection
         │                       │                        │
         └───────────────────────┼────────────────────────┘
                                 │
                                 ▼
                    Adaptive Sensor Fusion Engine
                                 │
                                 ▼
                    Drone / No Drone Classification
                                 │
                                 ▼
                        Threat Level & Alert System

---

🎯 Research Goals

1. Robust RF Signature Classification: Detect drone signals and identify flight modes under varying Signal-to-Noise Ratios (SNRs).
2. Acoustic Signature Detection: Implement lightweight, audio-spectrogram classifiers to detect drone rotors in the acoustic near-field.
3. Optimized Computer Vision: Deploy highly optimized, low-latency object classifiers/detectors for visual identification.
4. Adaptive Late Fusion: Fuse probabilities dynamically using entropy-based confidence weighting to manage sensor degradation.
5. Edge & TinyML Benchmarking: Benchmark latency, parameter sizes, RAM, and Flash memory consumption to target ARM Cortex microcontrollers.

---

📂 Repository Structure

The project directory tree is organized for research scale and reproducibility:

Vardan/
├── pyproject.toml              # Build system, tool configurations, and metadata
├── requirements.txt            # Python dependencies (PyTorch, Librosa, etc.)
├── environment.yml             # Conda environment definition
├── LICENSE                     # MIT License
├── configs/                    # Hyperparameters and options (No hardcoding)
│   ├── dataset.yaml            # Dataset details, sampling rates, splits
│   ├── preprocessing.yaml      # Signal windows, FFT size, wavelets
│   ├── training.yaml           # Learning rates, scheduler, epochs
│   ├── evaluation.yaml         # Decision thresholds, TinyML constraints
│   └── logging.yaml            # Rotating log configurations, ML tracking
├── data/                       # Data separation (Raw data is read-only)
│   ├── raw/DroneRF/            # Landing directory for official raw data
│   ├── interim/                # Temporary files during preprocessing
│   └── processed/              # Split/transformed datasets (FFT, spectrograms)
├── notebooks/                  # Single-responsibility research notebooks
│   ├── 01_DroneRF_Dataset_Exploration.ipynb
│   ├── 02_RF_Signal_Processing.ipynb
│   ├── 03_Baseline_Model_Reproduction.ipynb
│   ├── 04_Benchmarking.ipynb
│   ├── 05_Vardhan_RF_Module.ipynb
│   ├── 06_Acoustic_Module.ipynb
│   ├── 07_Vision_Module.ipynb
│   ├── 08_Multimodal_Fusion.ipynb
│   └── 09_Experiments.ipynb
├── src/                        # Modular python package
│   ├── data/                   # Data loader helpers
│   ├── preprocessing/          # Signal transforms (STFT, wavelets)
│   ├── features/               # Descriptive feature extraction
│   ├── visualization/          # Spectrogram and signal plotters
│   ├── models/                 # Baselines and VardhanRFNet architecture
│   ├── evaluation/             # Metrics calculation (Precision, Recall, F1)
│   ├── benchmark/              # Latency & memory profiling utilities
│   ├── fusion/                 # Adaptive late-fusion algorithms
│   └── utils/                  # Paths resolver and helpers (set_seed)
├── models/                     # Weights storage directory
│   ├── baselines/              # Trained weights for baseline models
│   ├── vardhan/                # Trained weights for Vardan models
│   └── checkpoints/            # Epoch-wise train checkpoints
├── experiments/                # Running tracking logs and results
│   ├── configs/                # Saved configs of individual runs
│   ├── logs/                   # System outputs / training logs
│   ├── runs/                   # Local logs / TensorBoard outputs
│   └── reports/                # Experiment summaries
├── results/                    # Research output exports (LaTeX/JSON)
│   ├── tables/                 # Text/LaTeX formatted results tables
│   ├── plots/                  # Static evaluation graphs
│   ├── metrics/                # Raw accuracy scores in JSON
│   └── confusion_matrices/     # Confusion matrix arrays in CSV/JSON
└── reports/                    # Literature notes and paper draft assets

---

⚙️ Installation

We support package management through conda or standard pip.

Method 1: Conda (Recommended for cross-platform stability)

conda env create -f environment.yml
conda activate vardan

Method 2: Pip

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

pip install -r requirements.txt

---

📡 Dataset Configuration

Important

The DroneRF dataset is NOT included in this repository due to licensing restrictions and file size limits. Users should download the raw data separately from the official source and extract the files into the data/raw/DroneRF folder.

Follow the guidelines inside data/README.md to set up raw signals correctly. Path loading must utilize src/utils/paths.py to resolve locations dynamically.

---

🔄 Research Workflow

To replicate research results:

1. Parse raw signals and explore data configurations using 01_DroneRF_Dataset_Exploration.ipynb.
2. Experiment with wavelets and STFT in 02_RF_Signal_Processing.ipynb.
3. Replicate baseline results using 03_Baseline_Model_Reproduction.ipynb.
4. Run edge latency benchmarking with 04_Benchmarking.ipynb.
5. Compare performance of the proposed Vardan model in 05_Vardhan_RF_Module.ipynb.

Ensure reproducible seeds are set via set_seed in src/utils/helpers.py.

---

📅 Milestones & Progress

Phase	Description	Status
Phase 1	Repository Scaffolding & Path Resolution	🟢 Completed
Phase 2	RF Signal Preprocessing & Baselines	🟡 In Progress
Phase 3	Acoustic Signal Classifier	⏳ Planned
Phase 4	Vision Object Detector	⏳ Planned
Phase 5	Adaptive Multimodal Sensor Fusion	⏳ Planned
Phase 6	Edge TinyML Optimization & Deployment	⏳ Planned

---

👨‍💻 Research Team

• Department of Computer Science & Engineering, Amrita Vishwa Vidyapeetham
• Team 53:
  • Adith Narayan G
  • S J Yuvan Dhurkesh
  • Subash Santhanam K
  • Sisthick S
• Project Guide:
  • Mr. Sumesh A K (Assistant Professor)

---

📑 Citation Placeholder

When publishing or referencing this project, please cite:

@article{vardan2026multimodal,
  title={Towards an Indigenous, Portable and Energy-Efficient Counter-UAS Detection System using Adaptive Multi-Modal Artificial Intelligence},
  author={Adith Narayan, G. and Yuvan Dhurkesh, S. J. and Subash Santhanam, K. and Sisthick, S. and Sumesh, A. K.},
  journal={arXiv preprint (Submission in Progress)},
  year={2026}
}

---

📜 License

This repository is released under the MIT License. See LICENSE for details.