Deep Learning-Based UAV-To-UAV Small Target Detection

In UAV-to-UAV detection, most of the target UAVs are small targets due to the changing viewpoints of the source UAVs and the unstable motion of the target UAVs. In order to improve the performance of UAV-to-UAV small target detection, we optimize its backbone network based on the YOLOv5 target detection algorithm by incorporating the channel-space self-attention mechanism to improve the attention to small target feature information. Meanwhile, we also propose a new loss function, Focal-CIoU, to make the network pay more attention to high-quality samples. In order to further improve the UAV localization accuracy, we also use the KMeans +  + algorithm to cluster the anchor frames in order to make the set anchor frames more compatible with the UAV target. In addition, we also use various data enhancement strategies such as mosaic, blend, copy, and paste to increase the richness of the samples. Experimental results show that our proposed algorithm performs very excellently in the tiny UAV target detection task. And achieves a performance of 81.6% and 41.9% on AP50 and AP{50:95} metrics, respectively, which are 7% and 4.6% better compared to the original YOLOv5 algorithm. Our proposed algorithm is also more competitive compared to the current SOTA algorithm.