A Small UAV Detection Method Based on Optical Flow and Visual Feature Fusion

The rapid growth of the unmanned aerial vehicle (UAV) market poses threats to public safety and personal privacy. Visual detection has become a cost-effective and important method for UAV detection, but it struggles with detecting small and distant UAVs that lack clear morphological features. To address this, we propose Fusion Net, a network detection approach that combines optical flow features with visual features for improved detection. Fusion Net utilizes Convolutional Neural Networks (CNNs) and Transformers for feature extraction and fusion, achieving excellent detection results. We also introduce Mask Augmentation, a new data augmentation method, to enhance network convergence and diversity in UAV scenes. Fusion Net, pre-trained on the Drone-vs-Bird dataset, exhibits excellent performance in detecting small drones, as demonstrated by its high detection accuracy in our self-made dataset evaluation.