A Training-time Friendly Network for Real-time Drone Detection

The rapid development of the unmanned aerial vehicle (UAV) market poses potential threats to public safety and personal privacy, and vision sensors are widely deployed to detect intruding drones due to the intuitiveness and accessibility of videos. However, the pixel area of long-distance intrusion drones is small and the morphological features are weak, which makes identification and positioning difficult. At present, a lot of work has focused on improving the detection accuracy of drones, but there is very little to improve the detection speed of drones and shorten the network training time. Therefore, we propose a drone detector with high accuracy, short training time, and fast inference speed. In this work, we choose Centernet from the consideration of a light-head, single-stage, and anchor-free design that enables fast inference. In addition, we improve the original framework of Centernet through the separation strategy of identification features and positioning features. Furthermore, encoding more training samples from annotated boxes has a similar effect to increasing the batch size, which helps to increase the learning rate and speed up the training process. To this end, we introduce a new method of Gaussian elliptic coding and propose a new loss function for it. Considering the complexity of background during flight and the small size of the drone compared to the background, a background filter is used to reduce the influence of the complex background. In this paper, the proposed model is trained and tested on the Drone-vs-Bird dataset, and its mAP reaches 83.7%, and its FPS reaches 29.6.