Ground-to-Air Low-Altitude Drone Detection System (G2A-LADD)

Note: The slight laser misalignment in the demonstration is due to unapplied hardware offset calibrations during recording, which can be dynamically fine-tuned via the control panel.

G2A-LADD is an industrial air defense prototype developed for the real-time detection, high-precision continuous tracking, and physical neutralization of low-altitude unmanned aerial vehicles (UAVs). The system processes the incoming video stream from a standard RGB camera at the artificial intelligence layer to detect targets, maintains tracking using predictive algorithms, and directs the turret mechanism to facilitate physical laser engagement.

System Capabilities

• Real-Time UAV Detection: High-precision detection against low-altitude micro and mini UAV targets powered by the YOLOv26s architecture.
• Predictive Target Tracking: Capability to maintain tracking by predicting the direction of motion even during instantaneous frame losses, thanks to the BoT-SORT tracking algorithm.
• Millisecond-Level Integration: Low-latency data communication via the PySerial protocol between the host computer and the ESP32 microcontroller.
• Dual-Mode Engagement: Operator-controlled manual ENGAGE or target-locked autonomous AUTO ENGAGE laser control.
• Dynamic Parameter Management: Instantaneous runtime modification of artificial intelligence detection thresholds and hardware calibrations via the desktop interface.

Hardware Layer and Power Architecture

The physical layer of the system is engineered with an external power architecture and industrial components to ensure absolute stability during high-torque servo movements.

Component	Function / Feature
Optical Sensor	Standard RGB Webcam providing the main video stream
Microcontroller	ESP32 Development Board facilitating data communication with the host system
Actuator Mechanism	Two-axis turret with Pan and Tilt movement capability powered by high-torque servo motors
Engagement Module	Industrial laser module focused on the target
Power & Driver Circuit	External 5V power supply, 2N2222 NPN transistor, 2200µF and 100nF capacitors, and current-limiting resistors

Technology Stack and Dependencies

• Development Environment: Python 3.12, Arduino IDE 2.x
• AI & Computer Vision: PyTorch, Ultralytics, OpenCV, NumPy
• User Interface: PySide6
• Data Management: SQLAlchemy

Development Environment Setup

git clone https://github.com/Melihg0/G2A_LADD.git

Method A: Conda Installation

conda env create -n G2ALADD -f environment.yml
conda activate G2ALADD

Method B: Pip Installation

python -m venv .venv
# For Windows:
.venv\Scripts\activate
# For Linux or Mac:
source .venv/bin/activate

pip install -r requirements.txt

Running the Application:

python main.py

ESP32 Firmware Flashing

1. Open the pan_tilt_controller.ino file under the firmware folder using Arduino IDE.
2. Install the ESP32Servo library via the Library Manager.
3. Select the ESP32 Dev Module board and complete the upload. Pin Configurations: Pan is mapped to GPIO 32, Tilt to GPIO 33, and Laser to GPIO 4.

User Interface and Control Panels

The control panel of the system is designed to allow the operator to manage all parameters without modifying the source code. Once started, you can control the system through the following panels:

• OPTICAL SENSOR and ENGAGE: Initiates the camera stream and enables autonomous or manual laser engagement on locked targets.
• AI Tuning Settings: Dedicated section for adjusting object detection confidence thresholds, target tracking memory duration, and turret movement optimizations.
• Hardware Config Settings: Calibration screen for ESP32 serial port selection, axis offset calibration, motor travel limits, and P-control gain adjustment.

Model Performance and Operational Metrics

The integrated deep learning model is trained on a high-resolution UAV dataset specifically compiled and optimized for this project. The system successfully filters out natural distractors such as clouds, light flares, or birds, delivering smooth detection and tracking performance on long-range targets.

The image below demonstrates the model's robustness in real-world field tests on unseen data, effectively maintaining tracking despite complex backgrounds and challenging targets.

Dataset / Model Parameter	Value / Ratio
Total Frames	~24,500
Positive Target Frames	~16,500
Negative Background Frames	~8,000
Operational Decision Confidence Threshold	0.46
Accuracy (mAP@50)	93.0%
Model Precision	95.6%
Model Recall	88.8%

Dataset and Model Weights Notice

The images in the custom dataset created for this project were compiled from licensed third-party platforms. Due to copyright restrictions, usage license limitations of the original sources, and the protection of intellectual property, the raw dataset and trained model weights cannot be shared publicly in this open-source repository.

3D Mechanical Production Files and Physical Setup

You can access the 5 industrial STL files consisting of the turret mechanism body, motor housing, and connection pins for computer-aided manufacturing via Printables or MakerWorld.