基于K-means和LCA的自动驾驶交通事故聚类分析

doi:10.16265/j.cnki.issn1003-3033.2025.07.1025

摘要/Abstract

摘要： 为了深入挖掘自动驾驶汽车(AV)道路交通事故的内在规律,仅依靠单一事故描述因素的统计分析是不够的,还需要进一步挖掘由多个因素相互作用所体现的综合潜在类别。鉴于AV事故数据既包含结构化信息,又包含叙事文本的特点,在类型识别过程中创新性地提出将K-means聚类分析与潜在类别分析(LCA)相结合的方法,首先,使用K-means方法从叙事文本中提取关键信息;然后,将其作为LCA模型的输入,克服LCA仅能利用现有事故报告中的结构化信息这一局限性;最后,采用美国加利福尼亚州的437起AV交通事故验证组合方法的有效性。结果表明:AV事故主要表现为4个综合类型;K-means 与 LCA 的组合方法能对含叙述文本的结构化信息实施高效的聚类分析。

关键词: K-means, 潜在类别分析(LCA), 自动驾驶, 聚类分析, 自动驾驶汽车(AV), 交通事故

Abstract:

To deeply explore the underlying patterns of road traffic accidents involving Autonomous Vehicles (AV), relying solely on the statistical analysis of individual accident description factors was insufficient. It was necessary to uncover further the comprehensive latent categories reflected by the interactions of multiple factors. Given that AV accident data contained structured information and narrative text, an innovative approach was proposed for type identification combining K-means clustering analysis and LCA. Specifically, the K-means method was used to extract key information from the narrative text, which was then fed into the LCA model to overcome the limitation of LCA being able to utilize only the structured information in existing accident reports. The effectiveness of this combined approach was verified using 437 AV traffic accidents in California, USA. The results show that AV accidents mainly manifest in four comprehensive types. The combined approach of K-means and LCA enables efficient clustering analysis of structured information that includes narrative text.

Key words: K-means, latent class analysis (LCA), autonomous driving, cluster analysis, autonomous vehicles (AV), traffic accidents

中图分类号:

X951交通运输安全

乔剑锋, 王亚楠, 吕淑然, 王汀, 夏学锋. 基于K-means和LCA的自动驾驶交通事故聚类分析[J]. 中国安全科学学报, 2025, 35(7): 192-200.

QIAO Jianfeng, WANG Yanan, LYU Shuran, WANG Ting, XIA Xuefeng. Cluster analysis of autonomous driving traffic accidents based on K-means and LCA[J]. China Safety Science Journal, 2025, 35(7): 192-200.

图/表 4

表1

表2

图1

表3

参考文献 20

[1]	赵晓华, 陈浩林, 李振龙, 等. 不同情景下自动驾驶接管行为的影响特征[J]. 中国公路学报, 2022, 35(9):195-214.
	ZHAO Xiaohua, CHEN Haolin, LI Zhenlong, et al. Influence characteristics of automated driving takeover behavior in different scenarios[J]. China Journal of Highway and Transport, 2022, 35(9):195-214.
[2]	郭延永, 刘佩, 袁泉, 等. 网联自动驾驶车辆道路交通安全研究综述[J]. 交通运输工程学报, 2023, 23(5):19-38.
	GUO Yanyong, LIU Pei, YUAN Quan, et al. Review on research of road traffic safety of connected and automated vehicles[J]. Journal of Traffic and Transportation Engineering, 2023, 23(5):19-38.
[3]	陈吉清, 翁楚滨, 兰凤崇. 智能车辆换道潜在冲突分析与风险量化方法[J]. 汽车工程, 2021, 43(11):1565-1576,1 586.
	CHEN Jiqing, WENG Chubin, LAN Fengchonog. Potential conflict analysis and risk quantification method of intelligent vehicle lane change[J]. Automotive Engineering, 2021, 43(11):1565-1576,1 586.
[4]	LIU Qian, WANG Xuesong, WU Xiangbin, et al. Crash comparison of autonomous and conventional vehicles using pre-crash scenario typology[J]. Accident Analysis and Prevention, 2021, 159:DOI:10.1016/j.aap.2021.106281.
[5]	ZHOU Rui, HUANG Helai, LEE J, et al. Identifying typical pre-crash scenarios based on in-depth crash data with deep embedded clustering for autonomous vehicle safety testing[J]. Accident Analysis and Prevention, 2023, 191:DOI:10.1016/j.aap.2023.107218.
[6]	DAS S, DUTTA A, TSAPAKIS I. Automated vehicle collisions in California: applying Bayesian latent class model[J]. IATSS Research, 2020, 44(4):300-308.
[7]	CHAE Q, KIM J, JANG J, et al. Development of risk-situation scenario for autonomous vehicles on expressway using topic modeling[J]. Journal of Advanced Transportation, 2022, 2022:1-19.
[8]	ALAMBEIGI H, MCDONALD A D, TANKASALA S R. Crash themes in automated vehicles: a topic modeling analysis of the California department of motor vehicles automated vehicle crash database[J]. arXiv: Applications, 2020: DOI:10.48550/arXiv.2001.11087.
[9]	GOODALL N J. Comparison of automated vehicle struck-from-behind crash rates with national rates using naturalistic data[J]. Accident Analysis and Prevention, 2023, 154:DOI:10.31224/osf.io/vbrx5.
[10]	杨文臣, 周燕宁, 田毕江, 等. 基于聚类分析和SVM的二级公路交通事故严重度预测[J]. 中国安全科学学报, 2022, 32(5):163-169.
	YANG Wenchen, ZHOU Yanning, TIAN Bijiang, et al. Traffic accident severity prediction for secondary highways based on cluster analysis and SVM model[J]. China Safety Science Journal, 2022, 32(5):163-169.
[11]	QIAO Jianfeng, LI Yu. Resource leveling using normalized entropy and relative entropy[J]. Automation in Construction, 2018, 87:263-272.
[12]	LIN Zijing, FAN Wei. Exploring bicyclist injury severity in bicycle-vehicle crashes using latent class clustering analysis and partial proportional odds models[J]. Journal of Safety Research, 2020, 76:101-117.
[13]	WELLER B E, BOWEN N K, FAUBERT S J. Latent class analysis: a guide to best practice[J]. Journal of Black Psychology, 2020, 46(4):287-311.
[14]	杨慧敏, 石琴, 陈一锴, 等. 基于混合聚类的农村公路单车事故影响因素分析[J]. 中国安全科学学报, 2020, 30(8):129-136.
	YANG Huimin, SHI Qin, CHEN Yikai, et al. Influencing factors for single vehicle accidents on rural highways based on hybrid clustering approach[J]. China Safety Science Journal, 2020, 30(8):129-136.
[15]	ARTHUR D, VASSIlVITSKII S. K-means++: the advantages of careful seeding[C]. Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, Society for Industrial and Applied Mathematics, 2007: 1027-1035.
[16]	QIAO Jianfeng, WANG Changfeng, GUAN Shuang, et al. Construction-accident narrative classification using shallow and deep learning[J]. Journal of Construction Engineering and Management, 2022, 148(9):DOI:10.1061/(asce)co.1943-7862.0002354.
[17]	NYLUND-GIBSON K, CHOI A Y. Ten frequently asked questions about latent class analysis[J]. Translational Issues in Psychological Science, 2018, 4(4):440-461.
[18]	罗崎瑞, 张道文, 周华, 等. 面向智能汽车预期功能安全的驾驶场景评价[J]. 中国安全科学学报, 2022, 32(8):140-145.
	LUO Qirui, ZHANG Daowen, ZHOU Hua, et al. Evaluation on driving scenarios for safety of intended functionality of intelligent vehicles[J]. China Safety Science Journal, 2022, 32(8):140-145.
[19]	YE Wanbao, WANG Chuanlin, CHEN Fuxiang, et al. Approaching autonomous driving with cautious optimism:analysis of road traffic injuries involving autonomous vehicles based on field test data[J]. Injury Prevention, 2020, 27(1):42-47.
[20]	SONG Yu, CHITTURI M, NOYCE D. Impact of event encoding and dissimilarity measures on traffic crash characterization based on sequence of events[J]. Accident Analysis and Prevention, 2023, 185:DOI:10.1016/j.aap.2023.107016.

类型变量	枚举实例	总数	占比/%
1 AV驾驶状态	行驶	251	57.4
1 AV驾驶状态	停止	186	42.6
2 AV驾驶模式	自动驾驶模式	252	57.7
2 AV驾驶模式	传统模式	185	42.3
3 AV碰撞前的状态	停车	176	40.3
	直行	123	28.1
	转弯	48	11
	倒车	21	4.8
	减速/减速停车	39	8.9
	变道	8	1.8
	停车/泊车	12	2.7
	其他行驶状态	10	2.3
4 AV碰撞类型	追尾	189	43.2
	侧刮	110	25.2
	侧撞	38	8.7
	正面碰撞	37	8.5
	撞物	37	8.5
	其他类型	26	5.9
5 AV损坏程度	未损坏	29	6.6
	轻度	332	76
	中度	65	14.9
	重度	11	2.5
6 CV碰撞前的状态	停车	13	3
	直行	188	43
	转弯	39	8.9
	倒车	27	6.2
	减速/减速停车	10	2.3
	变道	22	5
	停车/泊车	32	7.3
	超越其他车辆	17	3.9
	其他行驶状态	89	20.4
7 VRU	否	392	89.7
7 VRU	是	45	10.3
8人员受伤程度	未受伤	396	90.6
8人员受伤程度	受伤	41	9.4
9天气	晴朗	364	83.3
	多云	45	10.3
	雨天	26	5.9
	雾天	2	0.5
10照明	日光	293	67
	黎明	12	2.7
	路灯	127	29.1
	无路灯	5	1.1
11路面情况	干燥	405	92.7
11路面情况	潮湿	32	7.3
12道路情况	坑洞	2	0.5
	障碍物	3	0.7
	狭窄	4	0.9
	其他道路情况	10	2.3
	无异常	418	95.7

聚类1(203个样本)		聚类2(156个样本)		聚类3(79个样本)
特征词	TFIDF	特征词	TFIDF	特征词	TFIDF
车道(lane)	0.089 2	停止(stop)	0.103 4	停车(park)	0.194 9
右(right)	0.064 6	后部(rear)	0.099 4	侧面(side)	0.110 0
转弯(turn)	0.060 5	保险杠(bumper)	0.092 6	手动(manual)	0.073 1
大卡车(truck)	0.055 5	灯(light)	0.091 4	前部(front)	0.065 9
停止(stop)	0.051 0	交叉路口(intersection)	0.074 6	镜子(mirror)	0.063 4
前(front)	0.048 0	自动(autonomous)	0.067 2	倒车(reverse)	0.062 2
手动(manual)	0.047 0	在…之后(behind)	0.064 2	平行(parallel)	0.049 1
侧面(side)	0.044 1	红(red)	0.061 7	场地(lot)	0.047 5
后部(rear)	0.039 8	轻微(minor)	0.054 1	轻微(minor)	0.046 9
在…上(onto)	0.037 2	交通(traffic)	0.052 4	卷入(involve)	0.044 7
自动(autonomous)	0.037 0	靠近(approach)	0.043 5	后部(rear)	0.043 1
卷入(involve)	0.035 0	英里每小时(mph)	0.039 0	右(right)	0.042 9
街(boulevard)	0.033 0	转弯(turn)	0.038 1	自动(autonomous)	0.039 4
自行车(cyclist)	0.030 8	卷入(involve)	0.037 8	空间(space)	0.039 0
在…之后(behind)	0.030 5	在…之前(front)	0.036 1	马路牙子(curb)	0.037 9

类型变量	实例	聚类1	聚类2	聚类3	聚类4
1 AV驾驶状态	行驶	0	95.4	99	66.8
1 AV驾驶状态	停止	100	4.6	1	33.2
2 AV驾驶模式	自动驾驶模式	66.6	58.7	34.2	70.6
2 AV驾驶模式	传统模式	33.4	41.3	65.8	29.4
3 AV碰撞前的状态	停车	90.3	8.6	0	37.3
	直行	1	43.2	48	50.2
	转弯	1.6	21.7	10.6	8.3
	倒车	1.2	0	22.4	0
	减速/减速停车	0	20.5	6.6	4.2
	变道	0.6	4.4	0	0
	停车/泊车	5.3	0	3.5	0
	其他行驶状态	0	1.6	8.8	0
4 AV碰撞类型	追尾	59.7	44.6	11.7	29.4
	侧刮	22.7	25.4	30.9	20.8
	侧撞	8.2	13.3	0	12.5
	正面碰撞	7	10.2	5.8	16.6
	撞物	0	5.4	27.7	20.8
	其他类型	2.3	1.1	23.9	0
5 AV损坏程度	未损坏	5.9	5.2	11.5	4
	轻度	76.9	75.3	76.7	71
	中度	14.8	17.2	9	20.8
	重度	2.3	2.3	2.9	4.2
6 CV碰撞前的状态	停车	0	2.3	11	0
	直行	59.1	52.9	0	16.8
	转弯	5.7	16	0	16.6
	倒车	13	0	2.4	12.3
	减速/减速停车	4.7	1.3	0	0
	变道	2.9	8.9	0	12.4
	停车/泊车	1.2	0	34.1	4.4
	超越其他车辆	6.5	2.5	2.4	0
	其他行驶状态	7	16.2	50.1	37.4
7 VRU	否	89.7	88	90	100
7 VRU	是	10.3	12	10	0
8人员受伤程度	未受伤	88.2	88.9	96.1	100
8人员受伤程度	受伤	11.8	11.1	3.9	0
9天气	晴朗	91.3	90.2	78	0
	多云	6.8	8.6	22	5.2
	雨天	1.3	0.6	0	94.8
	雾天	0.6	0.6	0	0
10照明	日光	69.5	68.1	68.6	37.5
	黎明	2.9	1.3	5.8	0
	路灯	27.5	27.5	25.6	62.5
	无路灯	0	3.2	0	0
11路面情况	干燥	97	100	96.5	0
11路面情况	潮湿	3	0	3.5	100
12道路情况	坑洞	0	0	2.4	0
	障碍物	0	0	1.2	8.3
	狭窄	0.6	0.6	2.3	0
	其他道路情况	1.2	0	9.4	0
	无异常	98.2	99.4	84.7	91.7
13发生地点	十字路口/交叉路口	52.5	51.4	23.3	53.6
	直行道路	21.8	28.1	45.2	25.9
	停车场	3.5	1.3	14.1	8.2
	汇入汇出路口	6.4	17.1	0	12.3
	路边停车位	15.8	2.1	17.4	0