基于YOLO的街景影像中行人车辆检测方法  

The Method of Pedestrian and Vehicle Detection in Street View Image Based on YOLO

作　　者：王朝辉 王润哲 郭震冬 黄亮[1] WANG Zhaohui;WANG Runzhe;GUO Zhendong;HUANG Liang(Jiangsu Province Surveying and Mapping Engineering Institute, Nanjing Jiangsu 210013, China)

机构地区：[1]江苏省测绘工程院,江苏南京210013

出　　处：《北京测绘》2021年第11期1452-1455,共4页Beijing Surveying and Mapping

摘　　要：车辆等移动平台搭载全景相机、全球导航卫星系统接收机、惯性测量单元等传感器,可以采集360°的街景。在街景影像中检测行人车辆,一方面可以在数据发布时更好地保护隐私,另一方面可以为城市治理能力提升提供数据支撑。由于街景影像具有表达环境复杂、数据量庞大、目标距离变化大等特点,现有的行人车辆检测算法应用时均存在一定局限。为此,提出一种兼顾效率和精度的街景影像中行人车辆检测方法,利用球形投影原理和先验知识从街景的经纬映射图中划分出目标区域,接着使用你只观察一次(You Only Look Once,YOLO)v4模型从目标区域检测行人和车辆。实验证明,本文方法车辆检测的准确率高于91%,行人检测的准确率高于73%,每张平均耗时21 ms,具备准确率高、速度快的优势,满足实际项目的数据生产需求。The mobile platform,such as vehicles,equipped with panoramic camera,GNSS receiver,inertial measurement unit and other sensors,is able to collect 360°street view image.To detect pedestrian and vehicles in street view images,on the one hand,it can better protect privacy for data released,and on the other hand,it can provide data support for the improvement of urban governance capability.Because of the complex environment,large data and large change of target distance,the existing pedestrian and vehicle detection algorithms have some limitations in application.Therefore,in this paper,a method of pedestrian and vehicle detection was proposed in street view image which takes into account efficiency and accuracy.The target area was divided from the longitude and latitude map of street view by using the spherical projection principle and prior knowledge.Then,the pedestrian and vehicle were detected from the target area by using YOLO v4 model.The experiment showed that the accuracy of the method was better than 91%,the accuracy of pedestrian detection was better than 73%,and the average time for one image was about 21 ms.Our method had the advantages of high accuracy and fast speed,which met the data production demand of actual projects.

关 键 词：街景影像 经纬映射 你只观察一次(YOLO)模型 行人 车辆 

分 类 号：P228[天文地球—大地测量学与测量工程]