物联系统边缘端人流统计方法研究

Research on People Flow Statistics at the Edge of IoT System

针对传统人流统计方法效率低、对应用环境要求较高和难以满足多种需求场景的问题,本文将基于深度学习目标检测网络YOLOv5s与多目标跟踪算法Deep SORT融合起来,通过设置两个虚拟检测区域完成人流的方向检测与计数,实现了一种兼顾实时性与准确度的人流统计系统,可在智能物联系统的边缘端进行部署。经过测试,相比基于YOLOv3、YOLOv4以及YOLOv4-tiny的目标检测网络,本文方法的综合性能有明显提高,在特定环境下人流统计准确率最高接近100%。

The traditional people flow statistics method is inefficient,has high requirements on the application environment,and is difficult to meet the flexible and changeable demand scenarios.In view of this,this paper combines the deep learning-based target detection network YOLOv5s with the multi-target tracking algorithm Deep SORT,and completes the direction detection and counting of people flow by setting two virtual detection areas,realizes a real-time and accurate flow statistics system,it can be deployed at the edge of IoT system.Compared with the target detection network based on YOLOv3,YOLOv4 and YOLOv4-tiny,the comprehensive performance of this method is significantly improved,and the highest accuracy rate of people flow statistics in specific environment is close to 100%.