Advanced Engineering Informatics · Vol. 62 · 2024
Vision-Based Motion Prediction for Construction Workers Safety in Real-Time Multi-Camera System
Yuntae Jeon, Dai Quoc Tran, Almo Senja Kulinan, Taeheon Kim, Minsoo Park, Seunghee Park
We track construction workers and heavy equipment and predict their motion across multiple CCTV cameras by combining object detection, trajectory prediction-based tracking, and multi-camera association on a site map.
Highlights
- We suggest MPSORT (Motion Prediction Simple Online and Realtime Tracking) for a real-time hazard identification using Multi-CCTVs.
- By trajectory prediction-based tracking, we improve motion prediction accurarcy.
- In multi-camera systems, we localize objects on BEV map and identify hazards.
Abstract
Ensuring worker safety on dynamic construction sites is a significant challenge, especially as it is crucial to immediately identify potential hazards and warn workers. Existing computer vision-based motion prediction methods often overlook the false negative issue caused by the noisy environments of construction sites, and treat tracking and trajectory prediction as disconnected processes.
This study introduces MPSORT, a method that suggests trajectory prediction-based tracking with trajectory interpolation for vision-based automated safety monitoring. The proposed method predicts the future movements of construction workers and vehicles using multiple CCTV cameras, and localizes these predictions onto the construction site’s bird’s eye view (BEV) map. This enables to send the real-time warnings to workers in danger, preventing accidents such as collision, fall, and getting stuck.
We evaluated the performance of our method in both object tracking and trajectory prediction tasks on dataset from multiple CCTV cameras on construction sites. The object tracking results show an approximate 10% increase in the number of tracked objects, and the trajectory prediction results indicate an ADE of 7.138 and an FDE of 12.493, reflecting improvements of more than 5% and 2% in ADE and FDE, respectively, compared to previous methods. Overall, these findings are significant for minimizing accidents and enhancing safety and efficiency on construction sites.
Citation
@article{jeon2024vision,
title = {Vision-Based Motion Prediction for Construction Workers Safety in Real-Time Multi-Camera System},
author = {Jeon, Yuntae and Tran, Dai Quoc and Kulinan, Almo Senja and Kim, Taeheon and Park, Minsoo and Park, Seunghee},
journal = {Advanced Engineering Informatics},
volume = {62},
pages = {102898},
year = {2024},
doi = {10.1016/j.aei.2024.102898}
}