重庆理工大学学报(自然科学) ›› 2023, Vol. 37 ›› Issue (10): 247-254.

• 信息·计算机 • 上一篇    下一篇

基于级联多任务深度神经网络的施工现场车辆进出检测与识别算法

喻 捷,杨 倩,冯 欣   

  1. (1.重庆建工住宅建设有限公司,重庆 400000; 2.重庆理工大学 计算机科学与工程学院,重庆 400054; 3.重庆大学 大数据与软件学院,重庆 401331)
  • 出版日期:2023-11-20 发布日期:2023-11-20
  • 作者简介:喻捷,男,工程师,主要从事建筑施工管理与技术经济研究,Email:240586007@qq.com;通信作者 冯欣,女,博 士,副教授,主要从事计算机视觉研究,Email:xfeng@cqut.edu.cn。

Cascaded multi-task deep neural network based algorithm for construction site vehicle entry and exit detection and recognition

  • Online:2023-11-20 Published:2023-11-20

摘要: 施工场景下,对于进出车辆的管理涉及到建筑工地财产安全,以及对被污染车牌识 别不准确等问题。针对这些问题,提出了一种低算力的车辆进出检测、识别与追踪算法,大幅减 少了人力管理成本,实现了施工现场的车辆智能化管理。考虑到施工现场车型识别及车牌识别 涉及大小目标的混合检测,提出了一种级联多任务端到端的神经网络框架。利用 YOLO网络实 现了对施工现场车辆的检测与车型识别,在此基础上实现车牌定位和识别以及车辆目标跟踪。 车牌识别借助轻量级神经网络实现了端到端的精准识别,基于改进的 DeepSort目标跟踪算法实 现了对场内所有车辆的进出轨迹追踪。针对施工现场数据集缺乏问题,基于现有的门岗监控数 据,构建施工现场车辆检测数据集对多任务级联神经网络进行训练,并在 COCO2017数据集上 对算法进行了进一步验证,结果表明了算法的有效性与可靠性。

关键词: 车辆进出管理, 深度学习, 目标检测, 级联多任务

Abstract: In construction scenarios,the management of incoming and outgoing vehicles involves problems such as the safety of construction site property and inaccurate recognition of contaminated licence plates.In response to these problems,a low-computing power vehicle entry/exit detection,recognition and tracking algorithm is proposed,which significantly reduces the human management cost and realises the intelligent management of vehicles at construction sites.Considering that the tasks of model identification and licence plate recognition at construction sites involve mixed detection of large and small targets,this paper proposes a cascaded multi-task end-to-end neural network framework.Firstly,the YOLO network is used to achieve the detection of vehicles at the construction site and the recognition of vehicle models,and on this basis,the vehicle licence plate localisation and recognition,and vehicle target tracking are achieved at the same time.Licence plate recognition is achieved with the help of lightweight neural network to achieve end-to-end accurate recognition algorithm; target tracking algorithm based on improved DeepSort improves the tracking of all vehicles in and out of the site.For the lack of construction site dataset,based on the existing gate monitoring data,a construction site vehicle detection dataset is constructed to train the multi-task cascade neural network,and the algorithm is further validated on the COCO2017 dataset,and the implementation results show the effectiveness and reliability of the algorithm.

中图分类号: 

  • TP391