Unmanned aerial vehicle (UAV) vision: tiny object detection

Abstract

Unmanned Aerial Vehicle (UAV) vision has become an essential tool for a wide range of applications, including surveillance, traffic management, and agricul- tural monitoring. However, detecting tiny objects within aerial imagery remains a significant challenge due to factors like low resolution, scale variation, and non-uniform data distribution. In this dissertation, we introduce a novel frame- work, Adaptive Region-based Object Detection (AROD), specifically designed to enhance the detection performance of small objects in UAV images. Unlike conventional methods that rely on low-resolution feature maps for localization, AROD employs a high-resolution heatmap generator to improve the localization of tiny objects. Furthermore, we introduce the Local Density Module (LDM), which adaptively identifies and focuses on densely populated regions, thereby optimizing detection speed and accuracy. By replacing conventional regression loss with Gaussian Wasserstein Distance (GWD) loss, combined with L1 loss, AROD effectively targets small object detection while maintaining robustness across different scales. Comprehensive testing on two UAV datasets, VisDrone and UAVDT, shows that AROD outperforms existing top methods, especially in accurately detecting small objects. Future work aims to enhance feature learning and explore real-time detection in larger UAV scenes.