In the high-stakes environment of off-site construction, efficiency is paramount, but safety remains the foundation. As workers, heavy machinery, and massive structural panels move in tight, synchronized patterns, the risk of "struck-by" accidents—one of the industry's "Focus Four" hazards—is a critical concern. Researchers at the University of Alberta have introduced TrajGATFormer, a sophisticated AI framework designed to predict the future paths of both workers and moving obstacles, providing the necessary foresight for future collision-avoidance systems.
Executive Summary
TrajGATFormer is a hybrid deep learning framework that combines Graph Attention Networks (GAT) with Transformers to predict trajectories in dynamic construction environments. By modeling both social interactions and long-term movement patterns, it reduces trajectory prediction errors by up to 38% compared to traditional models like LSTMs.
Background and Problem Statement
Construction sites are inherently chaotic. Unlike robots in controlled factory settings, a construction worker’s movement is influenced by their specific tasks, their colleagues, and moving equipment. Traditional prediction methods often fall short in these environments:
- Knowledge-based models (e.g., Kalman Filters) often assume constant velocity, failing when a worker suddenly stops or turns.
- Standard Data-driven models (e.g., LSTMs) struggle to capture the "social" context—how a worker's path changes in response to a 2-ton floor panel swinging toward them.
A significant gap in existing research is that most models treat obstacles as stationary points. In reality, off-site construction involves moving components (panels, cranes) that fundamentally alter worker behavior. TrajGATFormer addresses this by integrating these dynamic obstacles into the prediction logic.
Methodology
The TrajGATFormer framework follows a comprehensive multi-stage pipeline to transform raw video input into accurate trajectory predictions.
- Detection & Tracking: The system utilizes YOLOv10n for real-time object detection and DeepSORT to maintain unique IDs for every worker and obstacle across video frames.
- Spatial Modeling (The GAT): The site is treated as a graph where workers and obstacles are nodes. A Graph Attention Network calculates the "influence" entities have on each other, effectively modeling social and environmental interactions.
- Temporal Modeling (The Transformer): While the GAT handles spatial relationships, the Transformer manages the temporal aspect. It uses an encoder-decoder architecture to process sequences of past movements and predict future coordinates over a 4.8-second horizon.
- Transfer Learning: Models were pre-trained on general pedestrian data (ETH dataset) to learn basic human motion before being fine-tuned on specific construction site data.
Two versions were developed: TrajGATFormer (focused on workers) and TrajGATFormer-Obstacle (which includes a dedicated encoder for moving panels).
Data and Experimental Setup
The research utilized two primary datasets:
- ETH Zurich Dataset: A standard benchmark for urban pedestrian trajectories used for pre-training.
- New Construction Dataset: A real-world dataset recorded at an off-site construction facility in Edmonton, Alberta. It features workers interacting with moving floor and roof panels. This dataset is unique because it captures "stop fractions"—the reality that construction workers spend significant time stationary while performing tasks, a behavior distinct from typical city pedestrians.
Results and Discussion
The proposed models significantly outperformed baselines like Social-LSTM and Social-GAN:
- Accuracy Boost: TrajGATFormer-Obstacle achieved an Average Displacement Error (ADE) of 1.03 meters and a Final Displacement Error (FDE) of 2.06 meters.
- Comparative Gains: This represents a 35% reduction in ADE and a 38% reduction in FDE compared to traditional LSTM-based approaches.
Key insights from the study include the realization that including moving obstacles in the model actually improved the accuracy of worker trajectory predictions. Understanding where a panel is moving helps the AI interpret why a worker is moving out of the way. Additionally, the hybrid GAT-Transformer approach successfully captures both "social force" (spatial) and "intent" (temporal).
Conclusion and Future Directions
While TrajGATFormer represents a major step forward, challenges remain regarding data scarcity for diverse moving obstacles and the model's difficulty with sudden, unpredictable maneuvers. Future work will focus on integrating these predictions into automated safety protocols, such as wearable haptic alerts or "kill switches" for machinery, to prevent collisions before they occur.
Summary updated