吉林大学学报(工学版) ›› 2021, Vol. 51 ›› Issue (3): 886-892.doi: 10.13229/j.cnki.jdxbgxb20200039

• 交通运输工程·土木工程 • 上一篇    下一篇

基于自动编码机-分类器的视频交通状态自动识别

彭博1,2(),张媛媛2,王玉婷2,唐聚2,谢济铭2   

  1. 1.重庆交通大学 山地城市交通系统与安全重庆市重点实验室,重庆 400074
    2.重庆交通大学 交通运输学院,重庆 400074
  • 收稿日期:2020-01-14 出版日期:2021-05-01 发布日期:2021-05-07
  • 作者简介:彭博(1986-),男,副教授,博士. 研究方向:交通视频、图像智能分析,城市干道交通状态识别与演变.E-mail:pengbo351@126.com
  • 基金资助:
    国家自然科学基金青年科学基金项目(61703064);重庆市科委基础前沿研究专项项目(cstc2017jcyjAX0473);重庆市技术创新与应用示范项目(cstc2018jscx-msybX0295);山地城市交通系统与安全重点实验室开放基金项目(2018TSSMC05)

Automatic traffic state recognition from videos based on auto⁃encoder and classifiers

Bo PENG1,2(),Yuan-yuan ZHANG2,Yu-ting WANG2,Ju TANG2,Ji-ming XIE2   

  1. 1.Chongqing Key Lab of Traffic System & Safety in Mountain Cities,Chongqing Jiaotong University,Chongqing 400074,China
    2.College of Traffic and Transportation,Chongqing Jiaotong University,Chongqing 400074,China
  • Received:2020-01-14 Online:2021-05-01 Published:2021-05-07

摘要:

为了及时、有效地识别道路交通状态,提出了结合自动编码机与分类器的视频交通状态识别方法。首先,建立交通状态视频图像数据集,对自动编码机隐藏层和降维数据维度等结构参数进行优化测试。然后,提出自动编码机定量评价方法,选出最优自动编码机模型A*。最后,将A*与线性分类器、支持向量机、深度神经网络、DNN Linear分类方法相结合,构建了4个交通状态识别模型。对前述模型及AlexNet、LeNet、GoogLeNet、VGG16等CNN模型进行训练测试,结果显示:本文模型精确率和召回率均为94.5%~97.1%,F1值均为94.4%~97.1%,CNN模型中AlexNet表现最佳,精准率、召回率以及F1值均为94%,表明A*与常用分类器结合,达到或超越了复杂CNN模型的交通状态识别效果。本文方法训练测试简便、计算成本低,适用于视频图像的交通状态识别。

关键词: 交通运输系统工程, 交通状态识别, 自动编码机, 交通视频, 深度学习

Abstract:

In order to recognize road traffic state timely and effectively, a method combining auto-encoder with classifiers was proposed for traffic state recognition from videos. Firstly, traffic video and image data sets were built, based on which auto-encoder structure parameter tests and optimizations were conducted over hidden layer and reduced data dimensionality. Then, a quantitative evaluation method for auto-encoders was put forward, thus the best auto-encoder model was presented as A*. At last, four traffic state recognition models were constructed by combining A* with Linear Classifier (LC), Support Vector Machine (SVM), Deep Neural Network (DNN) and DNN-LC respectively. The before mentioned models and CNN models including AlexNet, LeNet, GoogLeNet and VGG16 were trained and tested. Results show that, precision and recall of the proposed models are 94.5%~97.1%, and their F1 values are 94.4%~97.1%. Furthermore, AlexNet performs best among the four CNN models, with precision, recall and F1 value equal to 94%. Therefore, combiningA* and commonly used classifiers may reach or surpass the traffic state recognition effects of complicated CNN models. The proposed methods are convenient for training and testing with low computation cost, which are suitable for traffic state recognition from videos or images.

Key words: engineering of communications and transportation system, traffic status recognition, auto-encoder, traffic video, deep learning

中图分类号: 

  • U491.1

图1

本文方法流程"

图2

图像数据集制作"

图3

自动编码机结构"

表1

交叉试验模型结构"

输入数据维度降维数据维度隐藏层数量模型编号
32×323~6,步长为16~10,步长为21~12
64×643~6,步长为16~10,步长为213~24
128×1283~6,步长为16~10,步长为225~36

图4

交叉试验Lˉ0.01分布"

图5

损失曲线切线斜率kD分布"

表2

自动编码机备选模型结构参数"

模型编号输入数据维度编码器隐藏层数量降维数据维度编码器结构
32×3253(32×32,256)、(256,128)、(128,64)、(64,12)、(12,3)
32×3254(32×32,256)、(256,128)、(128,64)、(64,12)、(12,4)
32×3255(32×32,256)、(256,128)、(128,64)、(64,12)、(12,5)
32×3256(32×32,256)、(256,128)、(128,64)、(64,12)、(12,6)
64×6453(64×64,256)、(256,128)、(128,64)、(64,12)、(12,3)
64×6454(64×64,256)、(256,128)、(128,64)、(64,12)、(12,4)
64×6455(64×64,256)、(256,128)、(128,64)、(64,12)、(12,5)
64×6456(64×64,256)、(256,128)、(128,64)、(64,12)、(12,6)

表3

备选模型Lˉ0.005和斜率kD"

模型
Lˉ0.005/10-33.002.172.112.172.812.782.462.80
kD/10-7-1.95-1.45-1.38-1.42-1.80-1.78-1.57-1.77

表4

A*-分类器针对数据集A2各类状态的平均评价指标"

模型PˉRˉFˉ1
A*-SVM0.9620.9580.960
A*-DNN0.9620.9610.962
A*-Linear0.9450.9450.944
A*-DNN_ Linear0.9710.9710.971

图6

各模型对每类交通状态的分类效果"

图7

各模型畅通状态识别效果"

图8

各模型缓行状态识别效果"

图9

各模型拥堵状态识别效果"

表5

CNN模型测试效果"

模型名称模型结构精确率召回率F1训练测试时长/h
AlexNet5个卷积层、3个池化层、3个全连接层,1个分类层0.940.940.948.5
LeNet3个卷积层、2个池化层、1个全连接层,1个分类层0.820.620.718.3
GoogLeNet卷积层、池化层等22个网络层,1个分类层0.370.350.3616.8
VGG1613个卷积层、3个全连接层和5个池化层,1个分类层0.110.330.1716.2
A*-分类器5个编码隐藏层+分类器(线性分类、SVM、DNN等)0.95~0.970.95~0.970.94~0.972.3~2.7
1 Li L, Hu J, Huang Q, et al. A fuzzy hidden Markov model for traffic status classification based on video features[C]∥IEEE Conference on Computational Engineering in Systems Applications, Beijing, China, 2006: 2050-2055.
2 Quiros A R F, Bedruz R A, Uy A C, et al. Machine vision of traffic state estimation using fuzzy logic[C]∥IEEE Region 10 Conference, Marina Bay Sands, Singapore, 2016: 2104-2109.
3 崔华, 袁超, 魏泽发, 等. 利用FCM对静态图像进行交通状态识别[J]. 西安电子科技大学学报:自然科学版, 2017, 44(6): 85-90.
Cui Hua, Yuan Chao, Wei Ze-fa, et al. Traffic state recognition using state images and FCM[J]. Journal of Xidian University(Natural Science Edition), 2017, 44(6): 85-90.
4 Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[J]. Advances in Neural Information Processing Systems, 2012, 25(2): 1097-1105.
5 Ostrom Q T, Bauchet L, Davis F G, et al. The epidemiology of gliomain in adults: a "state of the science" review[J]. Neuro-Oncology, 2014, 16(7): 896-913.
6 Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]∥IEEE Conference on Computer Vision and Pattern Recognition(CVPR), Boston, USA, 2015: 1-9.
7 李映, 龚红丽, 梁佳熙, 等. 基于KSVD和PCA的SAR图像目标特征提取[J]. 吉林大学学报:工学版, 2010, 40(5): 1336-1339.
Li Ying, Gong Hong-li, Liang Jia-xi, et al. SAR image target feature extraction based on KSVD and PCA[J]. Journal of Jilin University(Engineering and Technology Edition), 2010, 40(5): 1336-1339.
8 侯阿临, 廖庆, 靳志娟, 等. 计算全息图的人工神经网络压缩算法[J]. 吉林大学学报:工学版, 2013, 43(): 21-24.
Hou A-lin, Liao Qing, Jin Zhi-juan, et al. Compression algorithm of computer-generated hologram based on artificial neural network[J]. Journal of Jilin University(Engineering and Technology Edition), 2013, 43(Sup.1): 21-24.
9 Gao S, Zhang Y, Jia K, et al. Single sample face recognition via learning deep supervised autoencoders[J]. IEEE Transactions on Information Forensics and Security, 2015, 10(10): 2108-2118.
10 Zhao C, Wan X, Zhao G, et al. Spectral-spatial classification of hyperspectral imagery based on stacked sparse autoencoder and random forest[J]. European Journal of Remote Sensing, 2017, 50(1): 47-63.
11 Ge P, Ren C X, Dai D Q, et al. Dual adversarial autoencoders for clustering[J]. IEEE Transactions on Neural Networks and Learning Systems, 2020, 31(4): 1417-1424.
12 Cheng H, Koc L, Harmsen J, et al. Wide & deep learning for recommender systems[C]∥Conference on Recommender Systems, New York, USA, 2016: 7-10.
[1] 宋震,李俊良,刘贵强. 基于深度学习和限幅模糊的变转速液压动力源恒流量预测方法[J]. 吉林大学学报(工学版), 2021, 51(3): 1106-1110.
[2] 袁哲明,袁鸿杰,言雨璇,李钎,刘双清,谭泗桥. 基于深度学习的轻量化田间昆虫识别及分类模型[J]. 吉林大学学报(工学版), 2021, 51(3): 1131-1139.
[3] 李锦青,周健,底晓强. 基于循环生成对抗网络的学习型光学图像加密方案[J]. 吉林大学学报(工学版), 2021, 51(3): 1060-1066.
[4] 王殿海,沈辛夷,罗小芹,金盛. 车均延误最小情况下的相位差优化方法[J]. 吉林大学学报(工学版), 2021, 51(2): 511-523.
[5] 张健,吴坤润,杨敏,冉斌. 智能网联环境下交叉口双环自适应控制模型[J]. 吉林大学学报(工学版), 2021, 51(2): 541-548.
[6] 宋现敏,张明业,李振建,王鑫,张亚南. 动态公交专用道的设置及其仿真分析评价[J]. 吉林大学学报(工学版), 2020, 50(5): 1677-1686.
[7] 赵宏伟,刘晓涵,张媛,范丽丽,龙曼丽,臧雪柏. 基于关键点注意力和通道注意力的服装分类算法[J]. 吉林大学学报(工学版), 2020, 50(5): 1765-1770.
[8] 谌华,郭伟,闫敬文,卓文浩,吴良斌. 基于深度学习的SAR图像道路识别新方法[J]. 吉林大学学报(工学版), 2020, 50(5): 1778-1787.
[9] 贾洪飞,丁心茹,杨丽丽. 城市潮汐车道优化设计的双层规划模型[J]. 吉林大学学报(工学版), 2020, 50(2): 535-542.
[10] 尹超英,邵春福,王晓全,熊志华. 考虑空间异质性的建成环境对通勤方式选择的影响[J]. 吉林大学学报(工学版), 2020, 50(2): 543-548.
[11] 郜峰利,陶敏,李雪妍,何昕,杨帆,王卓,宋俊峰,佟丹. 基于深度学习的CT影像脑卒中精准分割[J]. 吉林大学学报(工学版), 2020, 50(2): 678-684.
[12] 张大伟,祝海涛. 考虑行人差异性的人群疏散最优决策理论模型[J]. 吉林大学学报(工学版), 2020, 50(2): 549-556.
[13] 常玉林,袁才鸿,孙超,张鹏. 基于改进元胞传输模型的城市路网实际阻抗计算方法[J]. 吉林大学学报(工学版), 2020, 50(1): 132-139.
[14] 隋振,姜源. 基于MIMO类脑情感学习回路的横-纵向综合控制驾驶员模型[J]. 吉林大学学报(工学版), 2020, 50(1): 140-146.
[15] 宗长富,文龙,何磊. 基于欧几里得聚类算法的三维激光雷达障碍物检测技术[J]. 吉林大学学报(工学版), 2020, 50(1): 107-113.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
No Suggested Reading articles found!