VisDrone Object Detection

This project implements object detection on the VisDrone dataset using YOLOv10n, focusing on detecting people and vehicles in drone imagery.

🎯 Project Overview

The project uses the VisDrone dataset, which is a large-scale benchmark for visual object detection in drone imagery. We've implemented a simplified version focusing on two main object categories:

• People (combined from pedestrians and people classes)
• Vehicles (combined from cars, vans, trucks, buses, and motorcycles)

📊 Dataset

Original VisDrone Dataset Size

• Train: 6,471 images
• Validation: 548 images
• Test-dev: 1,610 images
• Test(challenge): 1,580 images

Reduced Dataset Size (for faster experimentation)

• Train: 500 images
• Validation: 100 images
• Test-dev: 100 images

Class Mapping

Original classes have been simplified into two main categories:

New Class	Original Classes
Person (0)	pedestrian (0), people (1)
Vehicle (1)	car (3), van (4), truck (5), bus (8), motor (9)

Note: bicycle (2), tricycle (6), and awning-tricycle (7) classes are ignored in this implementation.

🛠️ Tools & Technologies

• Deep Learning Framework: YOLOv10n
• Programming Language: Python
• Key Libraries:
  • PIL (Python Imaging Library)
  • tqdm (Progress bar)
  • os, pathlib (File operations)
  • random, shutil (Dataset reduction)

📁 Project Structure

visdrone-object-detection/
├── data/
│   └── VisDrone/
│       ├── VisDrone2019-DET-train/
│       ├── VisDrone2019-DET-val/
│       └── VisDrone2019-DET-test-dev/
├── preprocess/
│   ├── annotation_handler.py
│   └── data_reducer.py
└── notebooks/
    └── VisDrone_yolov10n_custom_label_e100_bs32.ipynb

🔧 Data Preprocessing

1. Dataset Reduction

The data_reducer.py script randomly samples a smaller subset of the original dataset while maintaining the corresponding annotations.

2. Annotation Conversion

The annotation_handler.py script converts VisDrone annotations to YOLO format and implements the class merging strategy:

• Converts box coordinates to YOLO format (normalized xywh)
• Merges multiple classes into two main categories
• Filters out ignored regions and unnecessary classes

🚀 Model Training

Training Configuration

path: /content/data  # dataset root dir
train: VisDrone2019-DET-train/images
val: VisDrone2019-DET-val/images
test: VisDrone2019-DET-test-dev/images

names:
  0: person
  1: vehicle

Training Experiments

1. Initial Run:

   • Model: YOLOv10n
   • Epochs: 10
   • Batch Size: 16

2. Extended Training:

   • Model: YOLOv10n
   • Epochs: 100
   • Batch Size: 32

📚 References

1. VisDrone Dataset: Vision Meets Drones: A Challenge
2. PaddleDetection Implementation: PaddleDetection GitHub

🤝 Contributing

Feel free to open issues or submit pull requests for improvements.

📄 License

[Add your license information here]

@article{zhu2021detection, title={Detection and tracking meet drones challenge}, author={Zhu, Pengfei and Wen, Longyin and Du, Dawei and Bian, Xiao and Fan, Heng and Hu, Qinghua and Ling, Haibin}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, volume={44}, number={11}, pages={7380--7399}, year={2021}, publisher={IEEE} }