RCSD-UAV: An object detection dataset for unmanned aerial vehicles in realistic complex scenarios

Abstract

Unmanned Aerial Vehicle (UAV) detection based on visible light plays an important role in urban low-altitude defense, public safety and other fields. However, the current dataset is limited by single scene, large object and other factors deviated from the actual application scene, making it difficult to meet the needs of sample-driven deep learning optical image UAV detection. Therefore, this paper proposed a novel realistic complex scenarios UAV object dataset (RCSD-UAV) to provide training data for UAV detection models based on artificial intelligence technology. All data were obtained from ordinary cameras or mobile phones in the real world, covering various commonly used UAV types and natural scenes. The dataset is classified according to the scene and the object size, and we evaluated several models and gave benchmarks. From the experimental results, it can be concluded that the detection of UAVs is challenging due to small size and complex background. The two-stage model has good detection effect but poor real-time performance. The one-stage model can better balance the detection effect and real-time performance.