Abstract: As the core technology of smart city construction, multi-object detection plays an irreplaceable role in establishing city map, mining city operation law and improving city operation efficiency. Based on the excellent target detection model YOLOv5, this paper selects typical scenes in cities, takes pedestrians and vehicles as detection targets, and improves the SPP module of the neck part of the original lightweight version algorithm to ensure detection speed and increase detection accuracy. Then, the improved algorithm and YOLOv5-s algorithm were deployed on the edge devices of smart cities at the same time, and a comparative test was conducted. Experimental results show that compared with YOLOv5-s algorithm, the improved algorithm in this paper can improve the multi-target detection accuracy by 15.7% when the speed remains basically unchanged, which solves the problem that the detection rate and accuracy of edge devices in smart cities cannot meet the requirements at the same time to a certain extent, and provides a theoretical reference and method basis for multi-target real-time detection in smart cities.

Key words: smart city, multi-target detection, edge equipment, real-time detection

摘要： 多目标检测作为智慧城市建设中的核心技术,在建立城市图谱、挖掘城市运行规律、提高城市运行效率中起着不可替代的作用。文中在优良的目标检测模型YOLOv5基础上,选取城市中的典型场景,以行人和车辆为检测目标,将原有轻量化版本算法neck部分的SPP模块进行改进,在保证检测速率的同时增加检测的精度,然后将文中改进算法和YOLOv5-s算法同时部署在智慧城市边缘设备上进行对比试验。试验结果表明,文中改进算法相较于YOLOv5-s算法在速度保持基本不变的情况下检测多目标精度提高6.8%,在一定程度上解决了智慧城市边缘设备检测速率和精度不能同时达到要求的问题,为智慧城市多目标实时检测提供了理论参考与方法依据。

关键词: 智慧城市, 多目标检测, 边缘设备, 实时检测