融合梯度改进YOLO和KCF模型的无人机目标识别跟踪算法  

Fusion gradient improved YOLO and KCF models for UAV target recognition and tracking algorithm

作　　者：王文胜 何君尧 黄民[1,2] 吴国新[2] Wang Wensheng;He Junyao;Huang Min;Wu Guoxin(College of Mechanical and Electrical Engineering,Beijing Information Science and Technology University,Beijing 100192,China;Key Laboratory of Modern Measurement and Control Technology,Ministry of Education,Beijing Information Science and Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学机电工程学院,北京100192 [2]北京信息科技大学现代测控技术教育部重点实验室,北京100192

出　　处：《仪器仪表学报》2025年第2期221-233,共13页Chinese Journal of Scientific Instrument

基　　金：国家重点研发计划(2020YFB1713203);北京市教育委员会科研计划(KM202411232023);北京信息科技大学“青年骨干教师”支持计划(YBT202403)项目资助。

摘　　要：针对无人机目标小、目标不显著的情况以及目标被遮挡后的再跟踪问题,提出一种将改进YOLO和改进KCF模型融合的无人机识别跟踪算法YOLO-G-KCF。该算法在特征处理方面将多通道梯度特征和原图特征通过特征级联的方式进行融合,并将融合特征引入YOLOv10算法之中,使改进算法对强光照、阴影等复杂光照条件下的目标有更好的检测效果;同时将多通道梯度特征信息引入KCF目标跟踪算法之中,通过设计一种多尺度特征检测,使KCF算法具有良好的尺度自适应;在头侧引入KCF跟踪结果,计算得新的损失函数Inner-IoU,更准确的识别跟踪目标。经实验表明,在由多个开源无人机视频目标跟踪组成的数据集上进行测试,YOLO-G-KCF算法取得95.3%的准确率;与YOLOv10原始模型相比,改进模型的mAP@0.5提高了1.37%,平均精度mAP@0.5达到了94.28%,且识别速度达到了112 FPS,能以100 FPS以上的速度运行,满足无人机目标识别跟踪的实时性需求。通过引入识别机制的跟踪并进行改进,在不损失速度的基础上,对比其他算法具有更好地识别跟踪效果。YOLO-G-KCF算法实现了对无人机在目标小、不显著以及遮挡后再跟踪等情况下的识别跟踪,识别准确率高、抗干扰能力强、硬件开发实时性好,具有一定的理论研究和工程应用价值。In view of the small and unobvious target of the UAV and the re-tracking problem after the target is occluded,an anti-UAV recognition and tracking algorithm,YOLO-G-KCF,which integrates improved YOLO and improved KCF models,is proposed.This algorithm introduces multi-channel gradient features and original image features into the YOLOv10 algorithm by means of feature concatenation in feature processing.Therefore,the improved algorithm has a better detection effect on targets under strong light,shadow,and other complex lighting conditions.Meanwhile,the multi-channel gradient features are introduced into the KCF target tracking algorithm,and a multi-scale detection is designed to make the KCF algorithm have good scale adaptability.The KCF tracking results are introduced after the Head,and the new loss function Inner-IoU is calculated to more accurately identify the tracking target.The experimental results show that the YOLO-G-KCF algorithm achieves a 95.3%accuracy rate when tested on the dataset comprising multiple open-source UAV video target tracking.This is in comparison with the original model of YOLOv10,wherein the improved model′s mAP@0.5 has an increase of 1.37%,and the average precision mAP@0.5 reaches 94.28%and the recognition speed reaches 112 FPS,which can operate at more than 100 FPS to satisfy the real-time requirements of UAV target recognition and tracking.Compared with other algorithms without sacrificing speed,introducing recognition mechanisms for tracking and improving them has better recognition and tracking effects.YOLO-G-KCF algorithm realizes the recognition and tracking of low-speed,small-sized,and low-altitude unmanned aerial vehicles in situations where the target is small,not prominent,and occluded.It has high recognition accuracy,strong anti-interference ability,good real-time hardware development,and certain theoretical research and engineering application value.

关 键 词：目标跟踪 目标识别 无人机跟踪 YOLOv10 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TH89[自动化与计算机技术—计算机科学与技术]