首页|基于计算机视觉技术的无人机检测跟踪方法

基于计算机视觉技术的无人机检测跟踪方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
针对无人机因目标较小而难以检测、检测速度慢、难于跟踪等问题,提出一种基于目标检测YOLOv5s算法和目标跟踪DeepSORT算法的无人机检测跟踪方法;采用自采数据集和公开数据集构建无人机检测数据集,使用针对小目标的数据增强方法以扩充数据集多样性;选择合适的YOLOv5算法模型实现无人机目标的精准、快速检测,引入基于批标准化层的模型剪枝方法进一步提高模型检测速度;利用DeepSORT算法实现无人机目标的实时追踪;通过对比YOLOv3、YOLOv4、Fast R-CNN以及改进前的YOLOv5算法,验证了本文方法在无人机检测方面的性能.结果表明:提出的无人机检测跟踪方法的全类平均精度达到0.947,每秒浮点运算次数达到2.93×109,在无人机检测的精度和速度方面均具有优势.
Unmanned Aerial Vehicle Detection and Tracking Method Based on Computer Vision Technology
To solve the problems of difficult detection,slow detection speed and difficult tracking for unmanned aerial vehicles(UAVs)due to small target,a UAVs detection and tracking method based on YOLOv5s algorithm and Deep-SORT algorithm was proposed.Self-collected data set and open data set were used to construct UAVs detection data set,and data enhancement method for small targets was used to expand the diversity of data set.The appropriate YOLOv5 al-gorithm model was selected to achieve accurate and fast detection of UAVs targets,and the model pruning method based on batch normalization layer was introduced to further improve the model detection speed.DeepSORT algorithm was ap-plied to realize the real-time tracking of UAVs targets.By comparing YOLOv3 YOLOv4,Fast R-CNN,and the unim-proved YOLOv5 algorithm have verified the performance of the proposed method in drone detection.The results show that the whole class average accuracy of the proposed UAVs detection and tracking method reaches 0.947,and the number of floating-point operations reaches 2.93×109 times per second,which has the advantages in detection accuracy and detec-tion speed of UAVs detection.

computer vision technologyunmanned aerial vehicle detectiontarget trackingmodel pruning

刘新锋、陈梦雅、李成龙、陈关忠、张晓

展开 >

山东建筑大学 计算机科学与技术学院,山东 济南 250101

山东建筑大学 行业大数据研究中心,山东 济南 250101

山东建筑大学 学报编辑部,山东 济南 250101

计算机视觉技术 无人机检测 目标跟踪 模型剪枝

国家自然科学基金项目山东省重点研发计划(重大科技创新工程)项目山东省自然科学基金山东建筑大学博士基金资助项目山东建筑大学博士基金资助项目

519753322021CXGC011204ZR2020QF029X19023Z0101XNBS20117

2024

10.13349/j.cnki.jdxbn.20240605.002

济南大学学报(自然科学版)
济南大学

济南大学学报(自然科学版)

CSTPCD北大核心
影响因子：0.441
ISSN：1671-3559
年,卷(期)：2024.38(4)