Pedestrian Pre-Collision Pose Generation from Eyewitness Descriptions Without Visual Evidence

The pedestrian pre-collision pose is a key factor influencing the severity of pedestrian-vehicle injuries. Current methods for acquiring human poses still rely heavily on visual data, such as images or videos. However, traffic accidents are highly unpredictable and often lack sufficient visual evidence, making it difficult to reconstruct pedestrian pre-collision poses and limiting the accuracy of injury assessments. Moreover, general models trained on daily activity datasets perform poorly in high risk scenarios with unusual poses and show limited robustness under occlusion, blur, and data scarcity. To address these challenges, we construct the Pedestrian-Vehicle pre-Collision Pose Text (PVCP2) dataset and propose the Pedestrian Pre-collision poSe Generation (PPSG) framework, which generates pedestrian pre-collision poses from eyewitness oriented structured descriptions without visual records. The PVCP2 dataset employs VLM-assisted annotation and template-based semantic refinement to produce 10K+ fine-grained text-pose pairs spanning diverse pre-collision scenarios. The PPSG framework is built upon a conditional variational autoencoder architecture, where an LLM-driven text encoder and a pose prior encoder jointly learn cross-modal latent representations. During inference, the model takes structured eyewitness descriptions as the sole input to generate multiple SMPL-based 3D pre-collision poses, enabling effective user selection and expert analysis. Experiments demonstrate superior performance over existing methods in both text-pose alignment and pose accuracy, offering a scalable and interpretable solution for downstream traffic safety applications in the absence of visual evidence. Code and data are available at https://github.com/wmj142326/PPSG.

Authors

Institutions

Publication Details

Journal
IEEE Transactions on Image Processing
Published
2026-11-17
DOI
https://doi.org/10.1109/tip.2026.3752867
Primary Topic
Human Pose and Action Recognition
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Pedestrian Pre-Collision Pose Generation from Eyewitness Descriptions Without Visual Evidence

Yu Meng, Xiaorui Peng, Zhongwei Qiu, Guangxuan Song et al.
1 links
IEEE Transactions on Image Processing
Human Pose and Action Recognition
article

Pedestrian Pre-Collision Pose Generation from Eyewitness Descriptions Without Visual Evidence

Yu Meng, Xiaorui Peng, Zhongwei Qiu, Guangxuan Song, Ying He, Meijun Wang, Chao Zheng
article en

Abstract

The pedestrian pre-collision pose is a key factor influencing the severity of pedestrian-vehicle injuries. Current methods for acquiring human poses still rely heavily on visual data, such as images or videos. However, traffic accidents are highly unpredictable and often lack sufficient visual evidence, making it difficult to reconstruct pedestrian pre-collision poses and limiting the accuracy of injury assessments. Moreover, general models trained on daily activity datasets perform poorly in high risk scenarios with unusual poses and show limited robustness under occlusion, blur, and data scarcity. To address these challenges, we construct the Pedestrian-Vehicle pre-Collision Pose Text (PVCP2) dataset and propose the Pedestrian Pre-collision poSe Generation (PPSG) framework, which generates pedestrian pre-collision poses from eyewitness oriented structured descriptions without visual records. The PVCP2 dataset employs VLM-assisted annotation and template-based semantic refinement to produce 10K+ fine-grained text-pose pairs spanning diverse pre-collision scenarios. The PPSG framework is built upon a conditional variational autoencoder architecture, where an LLM-driven text encoder and a pose prior encoder jointly learn cross-modal latent representations. During inference, the model takes structured eyewitness descriptions as the sole input to generate multiple SMPL-based 3D pre-collision poses, enabling effective user selection and expert analysis. Experiments demonstrate superior performance over existing methods in both text-pose alignment and pose accuracy, offering a scalable and interpretable solution for downstream traffic safety applications in the absence of visual evidence. Code and data are available at https://github.com/wmj142326/PPSG.

IEEE Transactions on Image ProcessingVol. PP
Nanyang Technological University (SG), Alibaba DAMO Academy (China) (CN), University of Science and Technology Beijing (CN)
China Scholarship Council
Human Pose and Action Recognition
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon