Anchor-Free Multi-UAV Detection and Classification Using Spectrogram

The advancements in unmanned aerial vehicle (UAV) technology have brought immense convenience to society. However, unauthorized UAVs pose a serious threat to personal privacy, public safety, and aviation security. Therefore, accurate UAV detection and classification are crucial. Moreover, with the increased popularity of UAVs, the likelihood of multiple UAVs appearing in the same area simultaneously has also dramatically increased. Recent studies demonstrate that object detectors, such as FasterRCNN and YOLO, can be used to detect and classify multiple UAVs based on spectrograms. To our best knowledge, the object detectors are directly used to classify UAV without considering the characteristics of the UAV signal spectrogram, which results in a decrease in recognition performance. In this paper, we analyze the characteristics of the UAV signal spectrogram in detail and conclude two problems, i.e., prior anchor mismatch and cross-domain detection, hindering the implementation of object detector for UAV recognition. To solve prior anchor mismatch, we propose an anchor-free detector based on keypoint and design a novel keypoints matching algorithm to improve recognition performance. To solve cross-domain detection, we propose an adversarial learning based data adaptation method, which can generate domain-independent and domain-aligned features. Finally, the experiments adopt practical spectrogram and synthetic spectrogram to verify the superiority of the proposed anchor-free detector and the effectiveness of the proposed data adaptation method.