文献类型：期刊文献

英文题名：Vision-Based Collision Prevention on Construction Sites: Integrating Trajectory Prediction and Uncertainty Modeling

作者：Wang, Zeli[1];Yan, Xuzhong[2];Liu, Qinyuan[3];Ma, Xiaoyan[3];Li, Heng[4]

机构：[1]East China Univ Sci & Technol, Dept Management Sci & Engn, Shanghai 200030, Peoples R China;[2]Zhejiang Univ Technol, Sch Management, Hangzhou 310014, Peoples R China;[3]Tongji Univ, Sch Comp Sci & Technol, Shanghai 201804, Peoples R China;[4]Hong Kong Polytech Univ, Dept Bldg & Real Estate, Hong Kong 999077, Peoples R China

年份：2026

卷号：152

期号：1

外文期刊名：JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT

收录：;EI(收录号：20254319397741);WOS:【SCI-EXPANDED(收录号:WOS:001626263700027)】；

基金：We are thankful for the financial support of the following grants: National Nature Science Foundation of China (No. 72304098); National Natural Science Foundation of China (No. 72201247); and Shanghai Pujiang Program (No. 22PJC030).

语种：英文

外文关键词：Collision prevention; Trajectory prediction; Transformer; Computer vision

摘要：Collisions between workers and vehicles are a leading cause of fatalities on construction sites around the world. While automation technology has been extensively applied to various aspects of construction safety management, including collision prevention, significant inaccuracies in estimating safe distances between workers and vehicles, coupled with limited hazard prediction capabilities, continue to hinder efforts to effectively prevent collision accidents on construction sites. To address these challenges, this study proposes a novel collision prevention method that integrates computer vision and trajectory prediction technologies. First, the You Only Look Once version 11 (YOLOv11) model and simple online and realtime tracking (SORT) algorithm were employed to accurately detect and track workers and vehicles, effectively extracting their movement trajectories in dynamic construction scenarios. Second, a transformer-based trajectory prediction algorithm was developed, enabling high-precision motion forecasting and providing critical data for risk region definition and collision warning. Finally, a dynamic risk region division method was designed, incorporating motion states and positional uncertainty to identify potential collision risks in real time. Experimental results demonstrated that the proposed system achieved over 90% accuracy in detecting workers and vehicles and successfully extracted their trajectories. The transformer-based prediction algorithm showed excellent short-term accuracy, providing reliable trajectory forecasts. Furthermore, the risk region definition method effectively identified potential collision areas, including those occurring during vehicle turning maneuvers. In summary, the proposed method performs well in complex scenarios, accurately predicting risk regions, including collision risks during vehicle turning maneuvers. This study provides a reliable and efficient solution for proactive collision prevention, offering robust technical support for improving safety management and reducing accidents in construction environments.