基于无人机视角的地面多军事目标跟踪算法

Ground Multi-Military Target Tracking Algorithm Based on Perspective of UAV

下载PDF

导出

摘要针对当前多目标跟踪算法在军事无人机视角的战场感知中存在轨迹中断和ID跳转问题,提出一种基于无人机视角的地面多军事目标跟踪算法VA-ByteTrack。首先,将EfficientFormerV2网络作为检测器YOLOX的主干网络,并在输出端添加卷积块注意力模块(Convolutional Bock Attention Module,CBAM),以提高检测器对小尺寸目标的特征提取能力,解决了小目标特征模糊导致的轨迹中断问题。其次,引入运动匹配代价交并比(Intersection Over,Union,IOU)和检测置信度S,构建了基于sigmoid函数的自适应卡尔曼滤波,以平衡高分检测框和低分检测框的不同协方差需求,解决了目标密集且相互遮挡所导致的频繁ID跳转问题。实验结果表明,相比SORT、DeepSORT、ByteTrack等主流算法,所提算法跟踪准确度分别提升1.7、1.4和1.0百分点,跟踪精度分别提升0.7、0.5和1.9百分点,轨迹中断问题分别降低49.4%、7.7%和7.2%,ID跳转问题分别降低46.3%、15.9%和12.1%,满足战场动态感知的实际需求。 Aiming at the problems of track interruption and ID jump in the battlefield perception from the perspective of military UAVs,this paper proposes a ground multi-target tracking algorithm VAByteTrack based on the perspective of UAVs.First,EfficientFormerV2 network was used as the backbone network of detector YOLOX,and CBAM attention mechanism was added to the output to improve the feature extraction capability of detector for small-size targets,and to solve the problem of track interruption caused by fuzzy features of small targets.Secondly,by introducing motion matching cost IOU and detection confidence S,an adaptive Kalman filter AKF based on sigmoid function is constructed to balance the different covariance requirements of high-score detection boxes and low-score detection boxes,and to solve the problem of frequent ID jump caused by dense targets and mutual occlusion.Experimental results show that compared with mainstream methods such as SORT,DeepSORT and ByteTrack,the tracking accuracy of the proposed method is increased by 1.7,1.4 and 1.0 percent point,the tracking accuracy is increased by 0.7,0.5 and 1.9 percent point,and the track interruption problem is reduced by 49.4%,7.7% and 7.2%,respectively.The ID jump problem is reduced by 46.3%,15.9% and 12.1% respectively,which meets the actual demand of battlefield dynamic perception.

作者马泽兵党长营曾志强李忠华李建素陈雪丽常文彪王文媛 MA Zebing;DANG Changying;ZENG Zhiqiang;LI Zhonghua;LI Jiansu;CHEN Xueli;CHANG Wenbiao;WANG Wenyuan(School of Mechanical Engineering,North University of China,Taiyuan 030051,China)

机构地区中北大学机械工程学院

出处《测试技术学报》 2025年第6期696-705,共10页 Journal of Test and Measurement Technology

基金山西省基础研究计划面上项目(202103021224199,20210302123047)。

关键词多军事目标跟踪无人机视角自适应卡尔曼滤波目标检测 multi-military target tracking perspective of UAV adaptive Kalman filtering object detection

分类号 TB96 [机械工程—光学工程]