基于编码-解码架构的高速公路交通冲突深度学习预测方法

魏伟; 郑来; 焦瀚生

doi:10.3963/j.jssn.1674-4861.2024.05.001

基于编码-解码架构的高速公路交通冲突深度学习预测方法

doi: 10.3963/j.jssn.1674-4861.2024.05.001

哈尔滨工业大学交通科学与工程学院哈尔滨 150090

基金项目:

国家自然科学基金项目 52072097

黑龙江省自然科学基金项目 LH2019E052

详细信息

作者简介:
魏伟（1997—），博士研究生. 研究方向：道路交通安全. E-mail：22b932011@stu.hit.edu.cn

通讯作者:
郑来（1985—），博士，教授. 研究方向：道路交通安全、智能交通. Email：zhenglai@hit.edu.cn

中图分类号: U491.31
计量
- 文章访问数: 177
- HTML全文浏览量: 57
- PDF下载量: 43
- 被引次数: 0
出版历程
- 收稿日期: 2023-12-06
- 网络出版日期: 2025-01-22

Deep Learning Prediction of Expressway Traffic Conflicts Based on The Encoder-Decoder Architecture

School of Transportation Science and Engineering, Harbin Institute of Technology, Harbin 150090, China

摘要

摘要: 为进一步揭示高速公路交通动态交通参数与交通冲突事件的关系。利用highD数据库以3 min为基本单位构建样本，共提取交通流和路段特征23个。基于交通冲突衡量指标后侵入时间（post encroachment time，PET）和不同阈值，统计车辆跟驰和变道过程中不同严重程度的冲突数量。基于随机森林回归（random forest regression，RFR）进行特征筛选，利用数值特征灰度图片转换技术（feature matrix to gray image，FM2GI）技术将各样本转化为灰度图片，结合2D-卷积提取图片特征，构建了卷积神经网络（convolutional neural networks，CNN），C-RFR，C-SVR这3个编码-解码模型，并与基准模型BP神经网络（back propagation neuralnetwork，BPNN）、RFR、支持向量回归（support vector regression，SVR）对比分析。结果表明，基于路段驶入车辆数和驶出车辆数等2个重要特征，平均车头时距、时间占有率、客车平均行驶速度、换道率和驶出点速度方差等5个有效特征，编码-解码架构下的CNN、C-RFR和C-SVR均优于直接应用基准模型，均方根误差（root mean squared error，RMSE）分别降低了12.6%，31.6%，18.5%，能够实现交通冲突的实时预测。其中CNN预测误差最小，且在应对不同严重程度的交通冲突预测时表现出良好的鲁棒性，及对2个关键参数表现出低敏感性。基于FM2GI技术和2D-卷积编码的CNN、C-RFR和C-SVR模型拓展了交通冲突预测建模的深度学习框架，可实现高速公路基本路段多严重程度交通冲突的可靠预测。
- 交通安全 /
- 交通冲突预测 /
- 编码-解码模型 /
- 深度学习 /
- 高速公路 /
- CNN
Abstract: The approach aims to uncover the relationship between dynamic traffic parameters and traffic conflict incidents and it further supports proactive safety control. The highD database is utilized to create sample data in 3-minute intervals, extracting 23 features related to traffic flow and road section characteristics. Based on the post encroachment time(PET)indicator, different thresholds are set to classify the severity of conflicts during car-following and lane-changing scenarios. The random forest regression(RFR)method is used to select the most critical features, while feature matrix to gray image(FM2GI)technology converts the sample data into grayscale images to enable 2D convolution to extract image features. Three encoder-decoder models, convolutional neural networks (CNN), C-RFR, and C-SVR are compared with baseline models (back propagation neural network (BPNN), RFR, and support vector regression(SVR). The results indicated that: based on two key features(the number of vehicles entering and exiting the road section)and five effective features(average headway, time occupancy, average driving speed of passenger cars, lane change rate, and variance of exit speeds), the CNN, C-RFR, and C-SVR models within the encoder-decoder framework outperformed the baseline models. Specifically, root mean squared error(RMSE)reduced by 12. 6%, 31. 6%, and 18. 5%, respectively, enabling real-time prediction of traffic conflicts. Among them, CNN exhibited the lowest prediction error and demonstrated strong robustness in predicting traffic conflicts of varying severities, along with low sensitivity to two key parameters. The CNN, C-RFR, and C-SVR models, utilizing FM2GI technology and 2D convolution encoding, expand the deep learning framework for traffic conflict prediction modeling, and achieve reliable predictions for multiple severities of highway traffic conflicts in basic road segments.
- traffic safety /
- traffic conflict prediction /
- encoder-decoder model /
- deep learning /
- expressway /
- CNN

HTML全文

图 1 highD采集地点分布

Figure 1. Locations of recordings included in highD

下载: 全尺寸图片幻灯片

图 2 交通冲突示意图

Figure 2. Traffic conflicts schematic diagram

下载: 全尺寸图片幻灯片

图 3 RFR示意图

Figure 3. Schematic diagram of RFR

下载: 全尺寸图片幻灯片

图 4 SVR原理示意图

Figure 4. Schematic diagram of SVR

下载: 全尺寸图片幻灯片

图 5 三层BPNN回归典型结构

Figure 5. Typical structure of a three-layer BPNN regression model

下载: 全尺寸图片幻灯片

图 6 数值特征到图像转换流程图

Figure 6. Flow chart of the transformation from numeric features to images

下载: 全尺寸图片幻灯片

图 7 2D-卷积层工作原理（过滤器：2×2×3）

Figure 7. Principle of 2D-convolution(Filter: 2 × 2×3)

下载: 全尺寸图片幻灯片

图 8 编码-解码模型框架结构示意

Figure 8. Framework of encoder-decoder model

下载: 全尺寸图片幻灯片

图 9 特征重要度排序

Figure 9. Ranking of feature importance

下载: 全尺寸图片幻灯片

图 10 模型预测值与观测值对比

Figure 10. Comparison between predicted value and observed value

下载: 全尺寸图片幻灯片

图 11 BP神经网络结构及训练过程

Figure 11. The structure and training process of BPNN

下载: 全尺寸图片幻灯片

图 12 图像转换示例图

Figure 12. Examples of transformed gray images

下载: 全尺寸图片幻灯片

图 13 卷积层结构示意图

Figure 13. Schematic diagram of Convolutional Layers

下载: 全尺寸图片幻灯片

图 14 全连接层结构

Figure 14. Structure of full connected layers

下载: 全尺寸图片幻灯片

图 15 编码-解码模型预测结果

Figure 15. Predictive results of encoder-decoder models

下载: 全尺寸图片幻灯片

图 16 模型评价指标对比图

Figure 16. Comparison of different model evaluation indicators

下载: 全尺寸图片幻灯片

图 17 基于PET阈值的鲁棒性分析

Figure 17. Robustness analysis based on different PET thresholds

下载: 全尺寸图片幻灯片

图 18 图像大小敏感性分析

Figure 18. Sensitivity analysis based on different image sizes

下载: 全尺寸图片幻灯片

图 19 不同批次大小敏感性分析

Figure 19. Sensitivity analysis based on different batch sizes

下载: 全尺寸图片幻灯片

图 20 模型训练过程对比

Figure 20. Comparison of different model training process

下载: 全尺寸图片幻灯片

表 1 highD数据库属性

Table 1. Main characteristics of highD

属性	描述	属性	描述
采集时间	2017年、2018年	车辆平均可见时长/s	13.6
采集地点	6个	车辆总数	110 000
总时长/h	16.5	小汽车数量	-90 000
总距离/km	45 000	货车数量	-20 000
平均时长/h	17	单向车道数	2~3
路段长度/m	400~420	变道行为/次	5 600

下载: 导出CSV

表 2 路段和交通流特征

Table 2. Characteristics of expressway segments and traffic flow

变量	描述	变量	描述
S	路段编号	V	观测时间内各车辆平均行驶速度的均值/（km/h）
D	上下行方向	V_C	观测时间内各客车平均行驶速度的均值/（km/h）
N	单向车道数	V_T	观测时间内各货车平均行驶速度的均值/（km/h）
i	限速情况（限速、不限速）	o	时间占有率/%
w	星期	H	平均车头时距/s
T	货车混入率	A_U	驶入观测路段处车辆点速度均值/（m/s）
Mc	路段长度与所有客车行驶时间均值之比/（km/h）	A_D	驶出观测路段处车辆点速度均值/（m/s）
M_T	路段长度与所有货车行驶时间均值之比/（km/h）	v_U	驶入观测路段处车辆点速度方差/（m/s）²
M	路段长度与所有车辆行驶时间均值之比/（km/h）	v_D	驶出观测路段处车辆点速度方差/（m/s）²
L	观测时间内车辆长度的均值/m	I	观测时间内驶入观测路段的车辆数/veh
W	观测时间内车辆宽度的均值/m	O	观测时间内驶出观测路段的车辆数/veh
R	观测时间内各车辆变道率/%

下载: 导出CSV

表 3 多严重程度交通冲突统计

Table 3. Statistics of different severities of traffic conflicts

冲突严重程度	冲突频数均值	冲突频数方差
PET≤1 s	68.53	31.47
PET≤1.5 s	106.02	47.11
PET≤2 s	132.36	55.98
PET≤2.5 s	149.74	61.00
PET≤3 s	163.55	64.14

下载: 导出CSV

表 4 基于不同核函数的SVR结果

Table 4. Results of SVR based on different kernels

模型	R²	MAE	RMSE
线性核函数	0.96	9.654	13.258
多项式核函数	0.88	13.649	22.795
径向基核函数	0.78	19.198	30.802

下载: 导出CSV

表 5 各模型评估结果

Table 5. Evaluation results of different models

模型	R²	MAE	RMSE
BPNN	0.981	6.243	9.055
SVR	0.96	9.654	13.258
RFR	0.973	7.111	10.742
CNN	0.985	5.559	7.913
C-SVR	0.981	6.287	9.072
C-RFR	0.982	6.572	8.75

下载: 导出CSV

参考文献(27)

[1]	ZHENG L, ISMAIL K, SAYED T, et al. Bivariate extreme value modeling for road safety estimation[J]. Accident Analysis & Prevention, 2018, 120: 83-91.
[2]	公安部交通管理局. 道路交通事故统计年报[R]. 北京: 公安部交通管理局, 2019. Ministry of Public Security, Transportation Bureau. The road traffic accidents statistics report in China[R]. Beijing: Ministry of Public Security, Transportation bureau, 2019. (in Chinese)
[3]	孟祥海, 张晓明, 郑来. 基于线形与交通状态的山区高速公路追尾事故预测[J]. 中国公路学报, 2012, 25(4): 113-118. doi: 10.3969/j.issn.1001-7372.2012.04.019 MENG X H, ZHANG X M, ZHENG L, Prediction of rear-end collision on mountainous expressway based on geometric alignment and traffic conditions[J]. China Journal of Highway and Transport, 2012, 25(4): 113-118. (in Chinese) doi: 10.3969/j.issn.1001-7372.2012.04.019
[4]	高雪林, 汤厚骏, 沈佳平, 等. 基于XGBoost的高速公路事故类型及严重程度预测方法[J]. 交通信息与安全, 2023, 41 (4): 55-63. doi: 10.3963/j.jssn.1674-4861.2023.04.006 GAO X L, TANG H J, SHEN J P, et al. A method for predicting the type and severity of freeway accidents based on XGBoost[J]. Journal of Transport Information and Safety, 2023, 41(4): 55-63. (in Chinese) doi: 10.3963/j.jssn.1674-4861.2023.04.006
[5]	马聪, 张生瑞, 马壮林, 等. 高速公路交通事故非线性负二项预测模型[J]. 中国公路学报, 2018, 31(11): 176-185. doi: 10.3969/j.issn.1001-7372.2018.11.019 MA C, ZHANG S R, MA Z L, et al. Nonlinear negative binomial regression model of expressway traffic accident frequency prediction[J]. China Journal of Highway and Transport, 2018, 31(11): 176-185. (in Chinese) doi: 10.3969/j.issn.1001-7372.2018.11.019
[6]	ZHENG L, SAYED T, MANNERING F. Modeling traffic conflicts for use in road safety analysis: a review of analytic methods and future directions[J]. Analytic Methods in Accident Research, 2021, 29: 100142. doi: 10.1016/j.amar.2020.100142
[7]	朱顺应, 蒋若曦, 王红, 等. 机动车交通冲突技术研究综述[J]. 中国公路学报, 2020, 33(2): 15-33. ZHU S Y, JIANG R X, WANG H, et al. Review of research on traffic conflict techniques[J]. China Journal of Highway and Transport, 2020, 33(2): 15-33. (in Chinese)
[8]	郑来, 侯芹忠, 郭延永, 等. 高速公路合流区冲突极值建模与交通事故预测[J]. 公路交通科技, 2022, 39(10): 132-140. doi: 10.3969/j.issn.1002-0268.2022.10.017 ZHENG L, HOU Q Z, GUO Y Y, et al. Conflict extremum modeling and traffic accident prediction for expressway merging areas[J]. Journal of Highway and Transportation Research and Development, 2022, 39(10): 132-140. (in Chinese) doi: 10.3969/j.issn.1002-0268.2022.10.017
[9]	郑来, 邓晓庆, 孟祥海. 基于PET极值统计的高速公路车道变换行为安全性研究[J]. 公路交通科技, 2016, 33(8): 120-126. ZHENG L, DENG X Q. MENG X H. Study on safety of lane changing behaviours on expressway based on PET extreme statistics[J]. Journal of Highway and Transportation Research and Development, 2016, 33(8): 120-126. (in Chinese)
[10]	孟祥海, 林兰平. 高速公路分合流区潜在事故风险研究[J]. 中国安全科学学报, 2015, 25(8): 164-170. MEGN X H, LIN L P. Research on potential crash risk in freeway merging and diverging areas[J]. China Safety Science Journal. 2015, 25(8): 164-170. (in Chinese)
[11]	马艳丽, 祁首铭, 吴昊天, 等. 基于PET算法的匝道合流区交通冲突识别模型[J]. 交通运输系统工程与信息, 2018, 18 (2): 142-148. MA Y L, QI S M, WU H T, et al., Traffic conflicts identification model based on post encroachment time algorithm in ramp merging area[J]. Journal of Transportation Systems Engineering and Information Technology, 2018, 18(2): 142-148. (in Chinese)
[12]	戢晓峰, 谢世坤, 覃文文, 等. 基于轨迹数据的山区危险性弯道路段交通事故风险动态预测[J]. 中国公路学报, 2022, 35(4): 277-285. doi: 10.3969/j.issn.1001-7372.2022.04.023 JI X F, XIE S K, QIN W W, et al. Dynamic prediction of traffic accident risk in risky curve sections based on vehicle trajectory data[J]. China Journal of Highway and Transport, 2022, 35(4): 277-285. (in Chinese) doi: 10.3969/j.issn.1001-7372.2022.04.023
[13]	SUN Y, MALLICK T, BALAPRAKASH P, et al. A data-centric weak supervised learning for highway traffic incident detection[J]. Accident Analysis & Prevention, 2022, 176: 106779.
[14]	ARVIN R, KHATTAK A J, QI H. Safety critical event prediction through unified analysis of driver and vehicle volatilities: Application of deep learning methods[J]. Accident Analysis & Prevention, 2021, 151: 105949.
[15]	ZHENG L, HU Z L, SAYED T. Traffic conflict prediction at signal cycle level using Bayesian optimized machine learning approaches[J]. Transportation Research Record: Journal of the Transportation Research Board, 2022, 2677(5): 183-195.
[16]	KRAJEWSKI R, BOCK J, KLOEKER L, et al. The highD dataset: a drone dataset of naturalistic vehicle trajectories on German highways for validation of highly automated driving systems[C]. 21^st International Conference on Intelligent Transportation Systems, Maui, HI, USA: IEEE, 2018.
[17]	ESSA M, SAYED T. Full Bayesian conflict-based models for real time safety evaluation of signalized intersections[J]. Accident Analysis & Prevention, 2019, 129: 367-381.
[18]	GROMPING, U. Variable importance assessment in regression: linear regression versus random forest[J]. The American Statistician, 2009, 63(4), 308-319.
[19]	翁剑成, 孙宇星, 孔宁, 等. 基于多源数据的公交专用道效能评价方法与影响模型[J]. 中国公路学报, 2022, 35(4): 267-276. WENG J C, SUN Y X, KONG N, et al. Evaluation method and influence model of bus lane performance based on multi-source data[J]. China Journal of Highway and Transport, 2022, 35(4): 267-276. (in Chinese)
[20]	司伟, 茆纬杰, 李宁, 等. 寒区沥青路面智慧化施工混合料温度预估机器学习模型[J]. 中国公路学报, 2023, 36(3): 81-97. SI W, MAO W J, LI N, et al. Research on machine learning model for temperature prediction of asphalt mixture in intelligent construction of asphalt pavement in cold regions[J]. China Journal of Highway and Transport, 2023, 36(3): 81-97. (in Chinese)
[21]	ZHENG Q K, XU C C, LIU P, et al. Investigating the predictability of crashes on different freeway segments using the real-time crash risk models[J]. Accident Analysis & Prevention, 2021, 159: 106213.
[22]	赵炜华, 刘浩学, 赵建有, 等. 基于BP神经网络的驾驶员昼夜动态空间距离判识规律[J]. 中国公路学报, 2010, 23(2): 92-98. ZHAO W H, LIU H X, ZHAO J Y, et al. Law of BP neural network-based space distance cognition of driver in dynamic environment at day and night[J]. China Journal of Highway and Transport, 2010, 23(2): 92-98. (in Chinese)
[23]	SHARMA A, VANS E, SHIGEMIZU D, et al. DeepInsight: a methodology to transform a non-image data to an image for convolution neural network architecture[J]. Scientific Reports, 2019, 9: 11399.
[24]	RAHIM M A, HASSAN H M. A deep learning based traffic crash severity prediction framework[J]. Accident Analysis & Prevention, 2021, 154: 106090.
[25]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.
[26]	LI P, ABDEL-ATY M, YUAN J H. Real-time crash risk prediction on arterials based on LSTM-CNN[J]. Accident Analysis & Prevention, 2020, 135: 105371.
[27]	WANG H, REN K, SONG J. A closer look at batch size in mini-batch training of deep auto-encoders[C]. 3rd IEEE International Conference on Computer and Communications, Chengdu, China: IEEE, 2017.