URBAN TRAFFIC FLOW ANALYSIS BASED ON DEEP LEARNING CAR DETECTION FROM CCTV IMAGE SERIES
URBAN TRAFFIC FLOW ANALYSIS BASED ON DEEP LEARNING CAR DETECTION FROM CCTV IMAGE SERIES
复制标题
DOI：
10.5194/isprs-archives-xlii-4-499-2018
复制
发表时间：
2018-09
期刊：
影响因子：
--
通讯作者：
M. Peppa;D. Bell;Tom Komar;Wen Xiao
中科院分区：
文献类型：
--
作者：
M. Peppa;D. Bell;Tom Komar;Wen Xiao
抽象。交通流分析是城市道路交通基础设施规划和管理的基础。自动车牌识别（ANPR）系统是用于车辆检测和行程时间估计的常规方法。然而，这样的系统特别集中在汽车牌照上，提供有限范围的道路使用者。开源深度学习卷积神经网络（CNN）的进步与免费提供的闭路电视（CCTV）数据集相结合，为各种道路使用者的检测和分类提供了机会。本文介绍的这项研究旨在通过对特定领域的低质量图像进行预训练的CNN模型进行微调来分析交通流模式，这些图像是在2018年的各种天气条件和季节中捕获的。这些图像是从英国泰恩河畔纽卡斯尔的东北联合管理局（NECA）旅行和运输数据收集的。结果表明，优化后的MobileNet模型具有98.2%的准确率、58.5%的召回率和73.4%的调和平均值，由于其快速的性能，可以用于大数据的真实的实时流量监测应用。与MobileNet相比，经过微调的Faster区域提案R-CNN模型提供了更好的调和平均值（80.4%）、召回率（68.8%）和更准确的汽车单位估计，可用于要求更高准确性的交通分析应用程序比速度。这项研究最终利用机器学习算法来更广泛地了解社会事件和极端天气条件下的交通拥堵和中断。
Abstract. Traffic flow analysis is fundamental for urban planning and management of road traffic infrastructure. Automatic number plate recognition (ANPR) systems are conventional methods for vehicle detection and travel times estimation. However, such systems are specifically focused on car plates, providing a limited extent of road users. The advance of open-source deep learning convolutional neural networks (CNN) in combination with freely-available closed-circuit television (CCTV) datasets have offered the opportunities for detection and classification of various road users. The research, presented here, aims to analyse traffic flow patterns through fine-tuning pre-trained CNN models on domain-specific low quality imagery, as captured in various weather conditions and seasons of the year 2018. Such imagery is collected from the North East Combined Authority (NECA) Travel and Transport Data, Newcastle upon Tyne, UK. Results show that the fine-tuned MobileNet model with 98.2 % precision, 58.5 % recall and 73.4 % harmonic mean could potentially be used for a real time traffic monitoring application with big data, due to its fast performance. Compared to MobileNet, the fine-tuned Faster region proposal R-CNN model, providing a better harmonic mean (80.4 %), recall (68.8 %) and more accurate estimations of car units, could be used for traffic analysis applications that demand higher accuracy than speed. This research ultimately exploits machine learning alogrithms for a wider understanding of traffic congestion and disruption under social events and extreme weather conditions.