气象2025，Vol.51Issue(4)：417-430,14.DOI:10.7519/j.issn.1000-0526.2025.010201
基于YOLOv5与DeepSort对天气雷达数据鸟杂波的识别与追踪
Identification and Tracking of Bird Clutter in Weather Radar Data Based on YOLOv5 and DeepSort
摘要
Abstract
According to the specific image feature that the bird echo shows obvious ring shape in the wea-ther radar reflectivity product,this article proposes an improved algorithm based on a lightweight convolu-tional neural network You Only Look Once Version5(YOLOv5)and multi-object tracking based on deep learning based simple online and realtime tracking(DeepSort).The training and test datasets are construc-ted based on radar volume scanning echo intensity.Data obtained from the Yingkou Weather Radar from 2020 to 2023.The bird echoes are tracked,respectively.Firstly,Shuffle Attention(SA),a lightweight attention mechanism,is introduced into YOLOv5 algorithm to improve the accuracy and effectiveness of the overall model checking.Secondly,in DeepSort algorithm,the original cross-merge-ratio intersection over union(IOU)matching mechanism is replaced by an improved loss function of object detection,distance-intersection over union(DIOU)matching mechanism.DIOU introduces the distance between the center points of the boundary box on the basis of calculating the overlap degree of the boundary box,so as to pro-vide more accurate positioning.The number of identification(ID)error matching and ID switching caused by partial occlusion overlap is reduced.The test results show that the optimized YOLOv5 algorithm im-proves the accuracy by 2.6 percentage point,the recall rate by 1 percentage point,and the average accura-cy of threshold values greater than 0.5 by 1.2 percentage point.The improved DeepSort algorithm reduces the number of ID switches by 2 times,and multi target tracking accuracy multi-object tracking accuracy(MOTA)increases by 4.5 percentage point,thus improves lightweight of the initial model.Generally,the overall checking performance is significantly improved,and may meet the actual demand for bird echo rec-ognition and tracking.关键词
深度学习/注意力机制/目标检测/目标追踪Key words
deep learning/attention mechanism/target detection/target tracking分类
天文与地球科学引用本文复制引用
姚文,李松书,王海江,张晶..基于YOLOv5与DeepSort对天气雷达数据鸟杂波的识别与追踪[J].气象,2025,51(4):417-430,14.基金项目
四川省科技厅重点研发项目(2023YFG0170)、中国气象局沈阳大气环境研究所和东北冷涡重点开放实验室联合开放基金项目(2023SYIAEKFMS08)和中国气象科学研究院基本科研业务费(2023Z019)共同资助 （2023YFG0170）
