Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (11): 319-327.DOI: 10.3778/j.issn.1002-8331.2302-0281

• Engineering and Applications • Previous Articles     Next Articles

Target Detecction and Real-Time Following for Unmanned Aerial Vehicle

LIU Rongqi, WANG Hongyu, HAN Jiaozhi   

  1. School of Electronic Information and Electrical Engineering, Shanghai Jiaotong University, Shanghai 200240, China
  • Online:2024-06-01 Published:2024-05-31

面向无人机目标的检测与实时跟随

刘瑢琦,王红雨,韩佼志   

  1. 上海交通大学 电子信息与电气工程学院,上海 200240

Abstract: With the increasing application of drones in various fields, the demand for drone regulation is also gradually increasing. At the same time, due to the limited computing power and energy of drone platforms, effective detection and following algorithms are particularly important. Currently, deep learning-based methods are very effective for target detection, but there are still issues with stability, safety and interference from target shadows when directly applied to aerial target tracking tasks. To address the problem of shadow interference during target detection, a shadow recognition algorithm based on the HSV color space is proposed, which can segment and identify the shadow areas of the detected object, thus eliminating the interference of shadows on target detection. In order to obtain more accurate 3D position of the target drone, a new positioning algorithm is designed, which combines the center point of the detection box with the relatively fixed center point of the target drone through weighted fusion, reducing the impact of target box size fluctuations on target position estimation. In the obstacle avoidance strategy, drone-related constraints are integrated to avoid excessive oscillation during drone following. Additionally, a dynamic self-localization algorithm is used to detect and correct the control results of the tracked drone in real-time, improving the robustness of the following task. The proposed algorithm is validated through simulation on the Unreal Engine 4 platform and physical experiments, which can keep the accuracy of drone following tasks at a level of 0.1 meters.

Key words: unmanned aerial vehicle (UAV) following, target detection, shadow recognition, center positioning, dynamic self positioning

摘要: 随着无人机在各个领域的应用越来越广泛,目前对于无人机的管制需求也逐步上升,同时由于无人机平台算力、能源有限,有效的检测与跟随算法显得尤为重要。基于深度学习的方法对于目标检测十分有效,但其直接应用于空中目标跟随这一任务还存在稳定性与安全性不足以及目标阴影的干扰这些问题。针对目标检测时阴影干扰问题,提出了基于HSV色彩空间的阴影识别算法,能够对检测对象阴影区域进行分割识别,从而排除阴影对目标检测的干扰;为了得到了更精准的目标无人机三维位置,设计了二次定位算法,将纯检测框中心点与目标无人机结构上相对固定中心点进行了加权融合,减少了目标框大小浮动对目标位置估计的影响;在避障策略中融合了无人机相关约束以此避免了无人机跟随时的过度震荡,并利用动态环境下的自定位算法对追踪无人机的控制结果进行检测与实时修正,提升整个动态跟随过程中的鲁棒性。所提算法经虚幻4平台下的仿真与实物实验中得到验证,能够将无人机跟随任务的跟随精度控制在0.1 m级。

关键词: 无人机跟随, 目标检测, 阴影辨识, 中心定位, 动态自定位