Fully automatic tracking method for UAVs based on YOLOv4 tiny

DOI:	10.11809/bqzbgcxb2023.09.005
Keywords:	UAV;aerial targets;YOLOv4 tiny;Kalman filter;embedded devices;fully automatic
Abstract:	UAVs equipped with vision systems have a wide range of applications in both civilian and military fields. In the actual military process, the use of machine vision to autonomously reconnaissance and strike small air targets is the key to improve the air combat performance of UAVs.An embedded target detection and tracking method mounted on quadrotor UAVs is proposed. The method is based on Jetson Xaiver NX airborne equipment and adopts YOLOv4 tiny as a deep learning architecture to train a specific aerial target model. Combined with binocular camera D435i, the method performs real time detection and 3D reconstruction of images acquired by UAVs. Aiming at the problem of short time occlusion, Kalman filter is adopted to help track the target. In addition, the UAV tracking control strategy module is combined to ensure that the quadrotor UAV can track aerial target objects autonomously and stably under various factors. According to the results of outdoor flight experiment, the feasibility and robustness of the fully automatic embedded system for target detection and tracking in complex environment are verified.
Issue:	Vol. 44 No. 9 (2023)
Published:	2023-09-28
PDF HTML