摘要
随着社会汽车数量的急剧增加,停车难已成为城市发展过程中的难题之一。针对这一问题,提出一种基于改进YOLO算法的停车场车位检测系统。通过网络获取、相机拍摄、三维建模3种方式分别获取不同场合、不同时间段及不同光照环境下的数据集。通过改变YOLOv3网络输入分辨率、增大动量、增大权值衰减值、增大批尺寸、减小抖动、选择合适的ignore_thresh值等6种不同措施对原网络进行改进,得到相对原网络检测效果更好的新网络YOLO-288×288。当IOU=50时,改进网络的平均检测精度由原来的86.5%提升到90.5%,查准率虽然降低了3%,但查全率提高了10%,平均IOU提高了1%。实验结果表明,新的YOLO-288×288网络优于原网络,其具有更高的检测精度、查全率及可靠性,能满足停车场的车位检测需求。
The number of cars in today’s society has increased dramatically and parking difficulties have become one of the problems in urban development.To solve this problem,a parking lot parking detection system based on the improved YOLO algorithm is proposed in this paper.The data sets in different occasions,different time periods and different illumination environments are obtained by three methods:network acquisition,camera shooting and three-dimensional modeling.By changing the input resolution of YOLOv3 network,increasing momentum,increasing weight attenuation,increasing batch size,reducing jitter and selecting appropriate ignore_thres value,a YOLO-288×288 network with better detection effect was obtained.When IOU=50,the average detection accuracy of the improved network was increased from 86.5%to 90.5%.Although the precision rate was reduced by 3%,the recall rate was increased by 10%and the average IOU was increased by 1%.The experimental results show that the new YOLO-288×288 network is superior to the original network,with higher detection accuracy,recall and reliability and can meet the detection requirements of parking spaces.
作者
陈俊松
何自芬
张印辉
CHEN Jun-song;HE Zi-fen;ZHANG Yin-hui(Faculty of Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China)
出处
《软件导刊》
2020年第4期37-41,共5页
Software Guide
基金
国家自然科学基金项目(61761024,61461022)。
关键词
YOLO网络
车位检测
机器视觉
卷积神经网络
YOLO network
parking space detection
machine vision
convolutional neural network
作者简介
陈俊松(1996-),男,昆明理工大学机电工程学院硕士研究生,研究方向为计算机视觉、深度学习;通讯作者:何自芬(1976-),女,博士,昆明理工大学机电工程学院副教授、硕士生导师,研究方向为图像处理;张印辉(1977-),男,博士,昆明理工大学机电工程学院教授、博士生导师,研究方向为图像处理、模式识别。
