A Density-Guided Regional Object Detection Method for Satellite Videos

Satellite Video Object Detection plays a crucial role in urban planning, traffic management, and military reconnaissance applications. However, detecting extremely small targets in satellite videos remains challenging due to low resolution and uneven target distribution across large coverage areas. In this paper, we present a novel method called DGROD—an improved CenterNet-based model designed to enhance the detection of tiny targets in satellite videos. First, we enhance feature maps by replacing the ResNet backbone with a High-Resolution Network and designing a feature enhancement neck module to integrate multi-scale features. Second, an unsupervised Local Scale Module (LSM) was introduced to detect and focus on regions where targets are clustered. Third, the loss function was modified by incorporating Shape-NWD loss, which combines Normalized Wasserstein Distance with shape considerations to improve small target detection performance. We validate our approach through experiments on the VISO dataset, where our model demonstrates significant improvements over the baseline in average precision (AP).