Sponsored by the National Research Foundation (NRF), South Korea through the Ministry of Education, Science, and Technology (MEST) under grant number 2018R1A6A1A03024003
Click to access dataset on IEEE Dataport
ABSTRACT
The deployment of unmanned aerial vehicles (UAV) for logistics and other civil purposes is consistently disrupting airspace security. Consequently, there is a scarcity of robust datasets for the development of real-time systems that can checkmate the incessant deployment of UAVs in carrying out criminal or terrorist activities. VisioDECT is a robust vision-based drone dataset for classifying, detecting, and countering unauthorized drone deployment using visual and electro-optical infrared detection technologies.
Dataset Methodology
The dataset consists of 20924 sample images and annotations from 6 drone models across 3 scenarios (cloudy, sunny, and evening), at different altitudes and distances (30m-100m), and in 3 different file formats (txt, xml, csv) that was generated at 12 different locations within a period of 1 year, 8 months by a team of domain experts.
The materials used for the data capturing includes drone models (Anafi-Extended, DJI FPV, DJI Phantom, EFT-E410S, Mavic2-Air, and Mavic2-Enterprise), drone controllers, mobile phone with controller application, high-definition digital cameras, and tripod stands. Each drone model was flown at different altitudes and distances at different times of day, week, and month. The video sequence of each scenario is recorded. Using reputable software applications, each video sequence is converted to JPEG image frames of 852 x 480 pixels and stored in repositories representing each model class and scenario sub-class. To minimize error, trained professionals carried out data cleaning on each repository by manually eliciting image frames without corresponding drones at the background. Data annotation was carried out by trained experts on each scenario sub-class in 3 file formats (txt, xml, and csv) by manually drawing bounding boxes around each image file to generate corresponding label files. To ensure consistency in naming convention and minimize error, each scenario sub-class label files are named to correspond to their image files and stored in repositories accordingly.
Dataset File Structure
The visioDECT dataset is arranged in 6 folders (representing the 6 drone models) with each folder having 2 sub-folders (representing the images folders and labels folders). Each image folder is made up of 3 scenario folders (representing cloudy, sunny, and evening) containing the image files stored in .JPG format. Each label folder contains 3 scenario annotation files (stored in .txt, .csv, and xml format) corresponding to the 3 scenario image folders. This makes it easy for classification, detection, and other image processing simulations on the dataset using developed models or different state-of-the-art artificial intelligence (AI) models.
