Drone identification based on CenterNet-TensorRT

Lei Tao, Tao Hong, Yichen Guo, Hangyu Chen, Jinmeng Zhang

Abstract: In recent years, due to the frequent occurrence of incidents of drone invasion and collisions, there exists a higher rate of safety accidents in crowded places. With the rapid development of drone recognition in the security field, there exist more and more drone recognition methods, most of which are costly and hard to implement. In the context of the rapid development of 5G technology, we propose a method utilizing the existing monitoring network to acquire data and deep learning approaches to detect drone targets to achieve drone recognizing, tracking and positioning. Our approach uses an improved CenterNet model to detect the presence of drones in the video frames. It also takes advantage of TensorRT to accelerate the inference of the trained model and abstract the spatio-temporal features of the drones in each video frame. A proportion integration differentiation (PID) framework is used to adjust the center of the camera and acquire the consequential actual coordinate to achieve positioning. We establish a largescale database, which includes 1097 images of drones in variant environments. Both the COCO2017 dataset and our drone dataset are used to test the original model and the accelerated model of resdcn101 and resdcn18, respectively. Extensive evaluations on both datasets demonstrate that the accelerated models with TensorRT can achieve a higher speed and a maintained detection precision.