AerialEye (YOLOv8 Aerial Patrol)

Test & validation

This model was tested on representative aerial shots (VisDrone test domain). You can reproduce the inference test with the sample images provided in the repository:

yolo predict model=aerialEye.pt source='sample_train_1.jpg'

Model Description

This is a custom trained YOLOv8 model optimized for aerial patrol and object detection tasks using RGB imagery. It is designed to be lightweight, efficient, and capable of real-time deployment on edge devices like UAS/Drones.

• Model Type: YOLOv8 (Object Detection)
• Task: Object Detection
• Architecture: Ultralytics YOLOv8
• Supported Classes: person, car, truck, motorcycle, bicycle

Intended Use

• Primary Use Case: Real-time aerial patrol, surveillance, traffic monitoring, and security.
• Deployment: Edge devices, cloud inference APIs, and ground-station software.

Limitations & "What it can't do"

To be completely transparent, here are the known limitations of this model in production environments:

• Angle Dependency: This model struggles with strictly ground-level perspectives (CCTV or dashcam angles) or perfectly top-down orthomosaic mapping views. It prefers a 30-60 degree slanted drone camera perspective.
• Low Light/Night conditions: As an RGB-trained model, performance degrades severely in low light or foggy conditions. It is not a substitute for a Thermal/IR trained model.
• Tiny Objects: While it detects vehicles well, predicting pedestrians or bicycles from extreme altitudes (>120 meters) will drop confidence significantly without an optical zoom.
• Class Limitation: Only objects conforming strictly to the 5 trained classes (person, car, truck, motorcycle, bicycle) will be detected. Other objects like buses, trains, or animals will be either ignored or misclassified.

Training Data

This model was trained on the VisDrone dataset (Vision Meets Drones), a large-scale aerial dataset captured by various drone platforms. The dataset contains rich annotations for vehicles, pedestrians, and cyclists across a variety of altitudes, weather conditions, and urban landscapes.

How to use

You can easily use this model with the ultralytics library.

Quick Setup Guide (Testing locally)

If you'd like to test the visual Gradio interface on your own machine:

# 1. Clone the repository
git clone https://huggingface.co/kilanisainikhil/AerialEye
cd AerialEye

# 2. Setup your python environment and install requirements
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# 3. Run the interactive UI
python app.py
# -> Open http://127.0.0.1:7860 to test your images or videos!

Downloading the Weights from Hugging Face

To download the model explicitly via Python into your own custom application, you can use the Hugging Face Hub library:

pip install huggingface_hub

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

# Download the PyTorch model
model_path = hf_hub_download(repo_id="kilanisainikhil/AerialEye", filename="aerialEye.pt")

# Load and run inference
model = YOLO(model_path)
results = model.predict("your_image.jpg")

(Alternatively, you can manually download aerialEye.pt and aerialEye.onnx from the Files tab of this repository).

Installation

pip install ultralytics opencv-python-headless

Python Inference

from ultralytics import YOLO
import cv2

# Load the custom model
model = YOLO('aerialEye.pt')

# Run inference on an image
results = model('sample.jpg')

# Display the results
for result in results:
    result.show() # to display
    # result.save(filename='result.jpg') # to save to disk

License

Provided under the MIT License. This model and inference code can be fully used in commercial, academic, or closed-source applications without aggressive copyleft restrictions.