基于单目姿态估计的行人碰撞损伤评估研究

doi:10.3969/j.issn.1674-8484.2026.01.003

摘要/Abstract

摘要：

为实现交通事故中行人应急姿态的量化损伤评估，该文提出一种基于单目姿态估计的车前行人碰撞损伤评估方法。利用姿态估计算法，从单视角事故图像中提取姿态参数，并通过关节匹配将蒙皮多人线性（SMPL）模型映射至多刚体模型，实现应急姿态的快速重建；结合MADYMO仿真软件，对站立、前倾、下蹲和跑步避让4种典型姿态开展股骨损伤的碰撞仿真。结果表明：该方法在仿真精度上与人工建模一致，4类姿态下的AIS2+与AIS3+损伤风险概率平均差值分别低至-3.5%和-1.22%；该方法提升了姿态建模的自动化与复现性，可为车辆结构优化和行人保护设计提供定量化参考。

关键词: 应急姿态, 姿态估计, 蒙皮多人线性（SMPL）模型, 多刚体模型, 损伤评估

Abstract:

A monocular pose-estimation-based method for evaluating pedestrian collision injuries in front of a vehicle was proposed to enable quantitative injury assessment of pedestrian emergency postures in traffic accidents. Posture parameters were extracted from single-view accident images using a pose estimation algorithm, and a skinned multi-person linear (SMPL) model was mapped to a multi-body model through joint matching to achieve rapid reconstruction of emergency postures. Subsequently, collision simulations of femur injuries were conducted using MADYMO under four typical postures, including standing, leaning forward, squatting, and evasive running. The results demonstrate that the proposed method achieves simulation accuracy comparable to that of manual modeling, with the mean differences in AIS2+ and AIS3+ injury-risk probabilities across the four postures as low as -3.5% and -1.22%, respectively, indicating that the proposed method improves the automation and reproducibility of posture modeling, and can provide quantitative references for vehicle structural optimization and pedestrian protection design.

Key words: emergency pose, pose estimation, skinned multi-person linear (SMPL) model, multi-body model, injury assessment

中图分类号:

U491.31

王美军, 孟宇, 郑超, 彭晓睿, 许焱. 基于单目姿态估计的行人碰撞损伤评估研究[J]. 汽车安全与节能学报, 2026, 17(1): 33-39.

WANG Meijun, MENG Yu, ZHENG Chao, PENG Xiaorui, XU Yan. Research on pedestrian collision injury assessment based on monocular pose estimation[J]. Journal of Automotive Safety and Energy, 2026, 17(1): 33-39.

图/表 8

参考文献 21

[1]	张道文, 刘奇, 邱健斌, 等. 人车碰撞中行人的应急姿态对行人的损伤影响[J]. 汽车安全与节能学报, 2021, 12(4): 456-466.
	ZHANG Daowen, LIU Qi, QIU Jianbing, et al. Influence of pedestrian’s emergency posture on pedestrian injury in vehicle-pedestrian collision[J]. *J Autom Safe Energ*, 2021, 12(4): 456-466. (in Chinese)
[2]	张鹏. 应急姿态对人车碰撞后行人运动形态和损伤研究[D]. 成都: 西华大学, 2019.
	ZHANG Peng. Study on pedestrian motion and injury after vehicle-pedestrian collision with emergency posture[D]. Chengdu: Xihua University, 2019. (in Chinese)
[3]	Badea-Romero A, Lenard J. Source of head injury for pedestrians and pedal cyclists: Striking vehicle or road?[J]. *Accid Analy Prev*, 2013, 50: 1140-1150.
[4]	SHI Shang, Otte D, LI Guibing, et al. Detailed assessment of pedestrian ground contact injuries observed from in-depth accident data[J]. *Accid Analy Prev, 2018, 110*: 9-17.
[5]	艾本清. 人车碰撞事故车速计算及行人损伤影响因素研究[D]. 西安: 长安大学, 2024.
	AI Benqing. Study on vehicle speed calculation in vehicle-pedestrian collisions and influencing factors of pedestrian injury[D]. Xi’an: Chang’an University, 2024. (in Chinese)
[6]	Crocetta G, Piantini S, Pierini M, et al. The influence of vehicle front-end design on pedestrian ground impact[J]. *Accid Analy Prev*, 2015, 79: 56-69.
[7]	Untaroiu C D, Meissner M U, Crandall J R, et al. Crash reconstruction of pedestrian accidents using optimization techniques[J]. *Int’l J Imp Engi*, 2009, 36(2): 210-219.
[8]	ZOU Tiefang, ZHA Aiming, LIU Qi, et al. Pedestrian gaits observed from actual pedestrian-vehicle collisions[J]. *Int’l J Crashworth*, 2022, 27(1): 1-23.
[9]	SHI Shang, Masson C, Llari M, et al. The predictive capacity of the MADYMO ellipsoid pedestrian model for pedestrian ground contact kinematics and injury evaluation[J]. *Accid Analy Prev, 2021, 149*: 105803.
[10]	尹均, 张道文, 肖凌云, 等. 人车事故中不同体型行人应急姿态仿真研究[J]. 中国安全科学学报, 2018, 28(2): 51-56. doi: 10.16265/j.cnki.issn1003-3033.2018.02.009
	YIN Jun, ZHANG Daowen, XIAO Lingyun, et al. Simulation study on emergency postures of pedestrians with different body types in vehicle-pedestrian accidents[J]. *Chin Safe Sci J*, 2018, 28(2): 51-56. (in Chinese)
[11]	LI Quan, SHANG Shi, PEI Xizhe, et al. Kinetic and kinematic features of pedestrian avoidance behavior in motor vehicle conflicts[J]. *Front Bioengi Biotech*, 2021, 9: 783003.
[12]	Schachner M, Schneider B, Klug C, et al. Extracting quantitative descriptions of pedestrian pre-crash postures from real-world accident videos[C]// 2020 IRCOBI Conf Proc. Farmington Hills: Int’l Res Coun Biomech Injur (IRCOBI), 2020: 231-249.
[13]	HUANG Yuan, ZHOU Qing, Koelper C, et al. Are riders of electric two-wheelers safer than bicyclists in collisions with motor vehicles?[J]. *Accid Analy Prev, 2020, 134*: 105336.
[14]	ZOU Tiefang, LIU Qi, ZHA Aiming, et al. New observations from real-world vehicle-pedestrian collisions in reducing ground related injury by controlling vehicle braking[J]. *Int’l J Crashworth*, 2022, 27(2): 614-631.
[15]	刘文虎. 人工智能在汽车自动驾驶中的应用[J]. 现代制造技术与装备, 2021, 57(8): 195-196.
	LIU Wenhu. Application of artificial intelligence in automobile autonomous driving[J]. *Mode Manufact Tech Equip*, 2021, 57(8): 195-196. (in Chinese)
[16]	张世豪. 基于深度学习的人体姿态估计方法研究[D]. 桂林: 桂林电子科技大学, 2024.
	ZHANG Shihao. Research on human pose estimation methods based on deep learning[D]. Guilin: Guilin University of Electronic Technology, 2024. (in Chinese)
[17]	Loper M, Mahmood N, Romero J, et al. SMPL: A skinned multi-person linear model[J]. *ACM Trans Graph (TOG)*, 2015, 34(6): 1-16.
[18]	Kocabas M, Athanasiou N, Black M J. Vibe: Video inference for human body pose and shape estimation[C]// Proc IEEE/CVF Conf Comput Visi Patt Recog. Seattle, WA, USA. 2020: 5253-5263.
[19]	刘卓异. 行人与厢式汽车碰撞事故中行人应急姿态对自身损伤影响研究[D]. 成都: 西华大学, 2020.
	LIU Zhuoyi. Study on the influence of pedestrian emergency posture on pedestrian injury in collisions with vans[D]. Chengdu: Xihua University, 2020. (in Chinese)
[20]	胡珊, 陈香, 黄皓, 等. 主被动安全融合场景下行人腿部碰撞损伤研究[J]. 汽车工程师, 2025(6): 16-26.
	HU Shan, CHEN Xiang, HUANG Hao, et al. Study on pedestrian leg injury in integrated active and passive safety scenarios[J]. *Autom Engineer*, 2025(6): 16-26. (in Chinese)
[21]	LIU Wu, BAO Qian, SUN Yu, et al. Recent advances of monocular 2D and 3D human pose estimation: A deep learning perspective[J]. *ACM Comput Surv*, 2022, 55(4): 1-41.

2D训练集	3D训练集	平均关节误差/mm	平均顶点误差/mm	加速度误差/ (mm·s^-2)
Insta	3DPW	87.32	103.81	25.18
Insta	MPII3D	97.31	116.01	27.55
PennAction	3DPW	93.27	109.71	30.34
PennAction	MPII3D	终止训练	终止训练	终止训练
Insta + PennAction	3DPW	89.36	106.20	27.08
Insta + PennAction	MPII3D	终止训练	终止训练	终止训练
Insta + PennAction	3DPW + MPII3D	91.67	108.17	27.81
Insta	3DPW + MPII3D	94.34	105.21	26.37
PennAction	3DPW + MPII3D	99.78	113.27	28.46

2D训练集	3D训练集	平均关节误差/mm	平均顶点误差/mm	加速度误差/ (mm·s^-2)
Insta	3DPW	87.32	103.81	25.18
Insta	MPII3D	97.31	116.01	27.55
PennAction	3DPW	93.27	109.71	30.34
PennAction	MPII3D	终止训练	终止训练	终止训练
Insta + PennAction	3DPW	89.36	106.20	27.08
Insta + PennAction	MPII3D	终止训练	终止训练	终止训练
Insta + PennAction	3DPW + MPII3D	91.67	108.17	27.81
Insta	3DPW + MPII3D	94.34	105.21	26.37
PennAction	3DPW + MPII3D	99.78	113.27	28.46

应急姿态	轴向载荷/kN		弯矩/Nm
应急姿态	基于姿态估计	基于手动构建	基于姿态估计	基于手动构建
站立姿态	6.11	6.12	403.22	393.15
下蹲姿态	9.03	10.21	340.91	342.29
前倾姿态	9.57	9.66	326.1	323.83
跑步姿态	6.02	5.99	397.29	377.32

应急姿态	轴向载荷/kN		弯矩/Nm
应急姿态	基于姿态估计	基于手动构建	基于姿态估计	基于手动构建
站立姿态	6.11	6.12	403.22	393.15
下蹲姿态	9.03	10.21	340.91	342.29
前倾姿态	9.57	9.66	326.1	323.83
跑步姿态	6.02	5.99	397.29	377.32

应急姿态	P(AIS2+) / %		P(AIS3+) / %
应急姿态	基于姿态估计	基于手动构建	基于姿态估计	基于手动构建
站立姿态	6.79	6.82	4.80	4.81
下蹲姿态	24.92	37.99	11.55	16.10
前倾姿态	30.53	31.53	13.47	13.82
跑步姿态	6.50	6.40	4.67	4.62
平均差值	-3.50		-1.22