Monitoring military vehicle depots

This paper presents a CNN-based approach for the localization and counting of closely parked military vehicles in Synthetic Aperture Radar (SAR) imagery. A significant challenge in this task is the separation of small, densely positioned vehicles with partially overlapping signatures. To address this, the problem is formulated as a segmentation task, and a U-Net model is trained using point-level annotations to predict vehicle locations. To mitigate the impact of class imbalance between vehicle positions and the background, the Tversky loss function is employed, which applies a greater penalty for missed detections of the underrepresented class. Finally, vehicle counts are derived from the predicted segmentation masks by counting connected components. To validate the approach, SAR simulation is used to generate a synthetic dataset comprising labeled SAR image data that depicts military vehicle depots from different perspectives. This dataset enables model training and allows for a comprehensive evaluation of the overall concept. The results demonstrate that the method successfully detects, separates and counts densely parked vehicles. The model trained on simulated data provides a suitable foundation for the subsequent exploitation of real image data in the future.