基于改进YOLO v7和DeepSORT的罐笼人员跟踪计数方法
Cage Personnel Counting Tracking Method Based on Improved YOLO v7 and DeepSORT
李敬兆 1刘敏 1郑鑫 2周小锋 2郎贵彬 2许志1
作者信息
- 1. 安徽理工大学 计算机科学与工程学院,安徽 淮南 232000
- 2. 淮沪煤电有限公司丁集煤矿,安徽 淮南 232000
- 折叠
摘要
煤矿副井罐笼载人时,超载现象时有发生,给副井运输安全带来了隐患.针对该问题,提出了基于改进YOLO v7 和DeepSORT的罐笼人员跟踪计数方法.首先利用煤矿副井罐笼区域的监控摄像头采集的视频制作罐笼行人数据集,再利用改进的YOLO v7 对罐笼人员进行识别:基于GhostNet网络改进YOLO v7 以提升模型的实时性;通过引入ACmix注意力机制提升模型对副井复杂背景人员的感知能力,并基于SIoU损失函数进一步增强模型的鲁棒性;最后利用优化后的DeepSORT算法对罐笼人员进行跟踪计数:基于CIOU优化DeepSORT算法的匹配准确度,使目标追踪更加稳定.试验结果表明:在构建的CP dataset数据集上,改进后的YOLO v7 网络的均值平均精度mAP达到了97.4%,算法的跟踪准确性MOTA和跟踪精度MOTP 分别达到了 95.74%和94.26%.
Abstract
When carrying people in auxiliary shaft,the phenomenon of overloading occurs frequently,which brings hidden danger to the safety of auxiliary shaft transportation.To solve this problem,a tracking and counting method based on improved YOLO v7 and DeepSORT is proposed in this paper.Firstly,the video collected by the surveillance camera in the cage area of the auxiliary shaft of the coal mine is used to make the cage pedestrian data set,and the improved YOLO v7 is used to identify the cage personnel:The YOLO v7 is improved based on GhostNet network to improve the real-time performance of the model;the ACmix attention mechanism is intro-duced to improve the perception ability of the model for people with complex background in auxiliary wells,and the robustness of the model is further enhanced based on the SIoU loss function.Finally,the optimized DeepSORT algorithm is used to track and count the cage personnel:The matching accuracy of the DeepSORT algorithm is optimized based on CIOU,which makes the target tracking more stable.The experimental results show that on the CP dataset constructed in this paper,the mean average precision mAP of the improved YOLO v7 network reaches 97.4%.The tracking accuracy(MOTA)and tracking accuracy(MOTP)of the proposed algorithm reach 95.74%and 94.26%,respectively.
关键词
煤矿罐笼安全/YOLO/v7/DeepSORT/GhostNet/目标检测与跟踪Key words
coal mine cage safety/YOLO v7/DeepSORT/GhostNet/target detection and tracking引用本文复制引用
出版年
2024
NETL
NSTL
维普
万方数据