Research on multi UAV target detection algorithm in the air based on improved CenterNet

In recent years, with the rapid development of UAV technology, the following problems and hidden dangers are increasingly prominent: “unauthorized flight” incidents occur frequently, security risks are serious, and social security is threatened. In order to quickly find UAV targets in these events and take relevant measures, this paper proposes a fast UAV target detection algorithm based on improved CenterNet, which can extract the depth features of the collected images, and use Gaussian kernel function to generate keypoint feature map, and output multiple target location information and category information. By modifying the structure and parameters of feature extraction network, the size of the model is greatly reduced and the running speed of the model is accelerated. At the same time, the detection effect of small target is improved obviously after adding DIoU loss. We train and test the algorithm on our own UAV dataset, and the experimental results achieve high accuracy (48.9 AP) and real-time detection speed (143fps). The improved algorithm is easy to be integrated into the mobile terminal or embedded system, providing support for the next step of alarming, jamming and drive away actions.