An Augmented Dataset for Vision-based Unmanned Aerial Vehicles Detection and Tracking

The rapid proliferation of Unmanned Aerial Vehicles (UAVs) or drones in military, disaster management, business, and entertainment applications has raised concerns about their potential airspace risks. Researchers are increasingly focused on developing methods for detecting and tracking UAVs with various data sources like radar, visual, acoustic, and radio-frequency data available. Among these, visual data stands out as cost-effective and amenable to analysis using Computer Vision (CV) techniques. However, vision-based tasks present challenges such as occlusions, shaky footage, and small UAVs at a distance, requiring timely and computationally efficient detection, especially given limited onboard computational power. To address these challenges, researchers are employing advanced Convolutional Neural Networks and Transformer-based deep learning models alongside classical image processing algorithms. An additional challenge is the scarcity of real-world UAV detection datasets, as existing ones often feature only similar types of UAVs and limited movement scenarios. Real-life situations involve various airborne objects, UAV movements can be erratic in any direction, and environmental conditions can be different. To overcome these limitations, we have proposed an augmented dataset that injects various drone types into existing UAV-to-UAV detection and tracking videos with different environmental conditions. Our dataset includes erratic movements of drones and other airborne objects like kites, and balloons, enhancing its realism and utility. We believe this dataset will alleviate the scarcity and constraints in the vision-based UAV-to-UAV detection and tracking research field. The augmented dataset is publicly available at https://github.com/hasiburrahman875/aud.