Recognition and Detection of UAV Based on Transfer Learning

With the increasing application scenarios of UAVs in industry, agriculture, military and other fields, the potential threats to national security and public security cannot be ignored. In addition, effective UAV detection and/or tracking is becoming an increasingly important security service. This paper integrates deep learning and image processing technology to conduct research in this context. In this paper, a transfer learning based UAV detection model (YOLOV5-UAV) is proposed. In order to reduce the influence of the amount of supervised data and the imbalance of target distribution on the performance of the model, the dataset is constructed based on self-shot videos and Internet downloaded videos in different natural scenes, combined with Mosaic data enhancement and adaptive scaling techniques. Therefore, the problem of data security is also effectively solved. Furthermore, real-time tests were carried out in two different time periods, namely day and night, from multiple scales, multiple perspectives and multiple natural scenes, for purpose of verifying the validity of the model. The applicability of different detection models is compared and analyzed for small target, moving background and weak contrast between UAV and background. The results show that YOLOV5-UAV model has a good performance in both detection accuracy and detection speed.