Mikiee/Person_Detection_and_Re-Identification_from_Low_Altitude_UAV-based_platform

Person Detection and Re-Identification from Low Altitude UAV-based Platform

Dataset Description

This dataset was collected as part of a master's thesis on person detection and re-identification using low-altitude UAV (drone) footage. It contains labeled aerial images captured from a DJI Mini drone, annotated in YOLOv8 format.

The dataset supports two tasks:

• Person Detection — detecting people in aerial drone footage
• Person Re-Identification (Re-ID) — recognizing and matching specific individuals across different frames and scenes

Dataset Structure

├── Detection/
│   ├── Batch-1.yolov8/
│   │   ├── train/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   ├── data.yaml
│   │   └── README.roboflow.txt
│   ├── Batch-2.yolov8/
│   ├── Batch-3.yolov8/
│   ├── Batch-4.yolov8/
│   └── Batch-5.yolov8/
├── Re-Identification/
│   ├── re-id-test1.yolov8/
│   │   ├── train/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   ├── data.yaml
│   │   └── README.roboflow.txt
│   ├── re-id-test2.yolov8/
│   │   ├── train/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   ├── data.yaml
│   │   └── README.roboflow.txt
│   ├── gallery_features_multiple_angles_misha.npy
│   └── gallery_features_multiple_angles_toghrul.npy
└── README.md

Detection Task

The detection subset contains 5 batches of labeled drone footage, each with approximately 100–300 images. Annotations are in YOLO format, with bounding boxes around detected persons.

Batches 3 and 5 correspond to "outside cage" scenarios and were used for a separate evaluation subset.

Classes

ID	Label
0	person

Re-Identification Task

The re-identification subset contains 2 batches of labeled images (re-id-test1 and re-id-test2), annotated with individual identity labels in YOLO format. This data was used to evaluate a person re-identification pipeline using the OSNet model.

Pre-computed gallery feature vectors are provided as .npy files (one per identity), enabling reproduction of the re-identification evaluation without re-extracting features.

Classes

ID	Label
0	misha
1	toghrul

Data Collection

• Platform: DJI Mini drone
• Altitude: Low altitude
• Annotation tool: Roboflow
• Annotation format: YOLO (TXT)
• Pre-processing/Augmentation: None applied to the exported dataset

Usage

Person Detection

from ultralytics import YOLO
 
model = YOLO("yolov8s.pt")
results = model.val(data="Detection/Batch-1.yolov8/data.yaml")

Person Re-Identification

import numpy as np
 
# Load pre-computed gallery features
gallery_misha = np.load("Re-Identification/gallery_features_multiple_angles_misha.npy")
gallery_toghrul = np.load("Re-Identification/gallery_features_multiple_angles_toghrul.npy")

License

This dataset is released under the CC-BY-4.0 license. You are free to share and adapt the data, provided appropriate credit is given.

Citation

If you use this dataset, please cite:

@mastersthesis{uav_person_detection_reid,
  title={Towards Autonomous Operation: A Perception and Localization Pipeline for DJI Consumer Drones},
  author={Mykhailo Donets, Toghrul Nasirov},
  year={2026},
  school={KU Leuven, faculty of Engineering Technology},
  note={Dataset available at https://huggingface.co/datasets/Mikiee/Person_Detection_and_Re-Identification_from_Low_Altitude_UAV-based_platform}
}