摘要
针对空间非合作目标(航天器、空间碎片等)因光线变化、目标旋转等导致跟踪失败问题,提出以带有C2F模块的YOLO(You Only Look Once)v8模型用于目标检测,并将经微调的全尺度网络(Omin-Scale Network,OSNet)替换StrongSORT中ResNet50进行重识别,以改进YOLO-StrongSORT算法。该算法首先以YOLOv8从采集的空间影像中检测目标;随即将检测结果分别送入包含OSNet网络的外性特征检测器和以NSA卡尔曼滤波器为核心的运动信息检测器提取检测目标的外观特征和预测目标运动轨迹;然后采用Vanilla匹配机制匹配连续帧中同一目标,实现对目标的跟踪;最后,用SNCOVT数据集和自研的近红外相机采集影像验证。结果表明,改进YOLO-StrongSORT算法在SNCOVT数据集上平均跟踪精度和成功率分别为61.7%和60.7%,较传统算法更准确;在实验室采集影像上跟踪精度和成功率分别为60.0%和56.2%,可用于空间目标的跟踪。
In order to solve the problem of tracking failure of space non-cooperative targets(spacecraft,space debris,etc.)due to variations in light conditions and target rotation,this paper proposes the use of the YOLOv8 model incorporating C2F module for target detection,and replaces ResNet50 in StrongSORT by the fine-tuned omni-scale network(OSNet)for re-identification,thereby improving the YOLO-StrongSORT algorithm.Firstly,YOLOv8 is used to detect the target from the collected spatial image.Then,the detection results are sent to the external feature detector containing the OSNet network and the motion information detector with the NSA Kalman filter as the core,respectively,to extract the appearance features of the detected target and predict its trajectory.Then,the Vanilla matching mechanism is used to match the same target in successive frames to realize its tracking.Finally,the SNCOVT dataset and the self-developed near-infrared camera are used to collect images for verification.The results show that the average tracking accuracy and success rate of the improved YOLOStrongSORT algorithm on the SNCOVT dataset are 61.7%and 60.7%,respectively,outperforming traditional algorithms.The tracking accuracy and success rate of the images collected in the laboratory are 60.0%and 56.2%,demonstrating suitability for space target tracking.
作者
李树清
陈雪旗
刘轶
夏鲁瑞
LI Shuqing;CHEN Xueqi;LIU Yi;XIA Lurui(Space Engineering University,Beijing 101416,China)
出处
《航天工程大学学报》
2025年第5期96-101,共6页
Journal of Space Engineering University
