摘要
由于目前空中“低慢小”目标存在无法快速、准确地检测和跟踪的问题,本文基于全景相机与变焦吊舱的多源信息融合方法,提出了一种检测跟踪方案。首先,基于YOLOv4-Tiny模型,本文在其颈部网络中引入坐标注意力机制,提出了YOLOv4-TCA算法,新算法所得的mAP较改进前提升2.3%。另外,本文利用间隔取帧手段,借助目标检测结果实时校正被跟踪目标尺度,改进得到AS-KCF目标跟踪算法。通过将上述算法应用到搭载全景相机和变焦吊舱的检测跟踪平台,实现了对空中“低慢小”目标的检测跟踪。实验结果表明,本方案目标重捕成功率较基线方案提升11.66%,目标重捕平均时间缩短2.6 s,具有一定的理论研究和工程应用价值。
With the growing number of airborne“low,slow and small”objects and the increasing related security problems,the first priority of the management strategy is to accurately detect and track such objects.However,“low,slow and small”objects in the air cannot be detected and tracked quickly and accurately,this paper proposes a detection and tracking scheme based on the multi-source information fusion method of panoramic camera and zoom pod.First,based on the YOLOv4-Tiny model,this paper introduces the coordinate attention mechanism in the neck network of the network model and proposes a YOLOv4-TCA algorithm,which improves the mAP of the model by 2.3%compared with that before the improvement.In addition,this paper improves the AS-KCF object tracking algorithm by using the interval frame taking method to correct the scale of the tracked object in real time with the help of object detection results.By applying the above algorithm to the detection and tracking platform equipped with panoramic camera and zoom pod,the detection and tracking of aerial“low,slow and small”objects is achieved.The experimental results show that the object recapture success rate of this scheme is 11.66%higher than that of the baseline scheme,and the average object recapture time is shortened by 2.6s,which has certain theoretical research and engineering application value.
作者
肖选杰
张浩天
艾剑良
XIAO Xuanjie;ZHANG Haotian;Ai Jianliang(Department of Aeronautics and Astronautics,Fudan University,Shanghai 200433,China)
出处
《复旦学报（自然科学版）》
CAS
CSCD
北大核心
2023年第5期605-614,共10页
Journal of Fudan University：Natural Science
