Trajectory Forecasting for Worker Safety in Construction Using Transformer and Graph Attention Networks

With the rise of residential housing demand worldwide, offsite construction emerges as a possible option to speed up construction while improving the safety of workers. However, offsite construction sites are normally a dynamic environment in which workers collaborate with various machinery and large moving objects, resulting in additional safety concerns. Accurate prediction of future trajectories is an important step in building a collision alarm system that can be utilized to mitigate such safety risks. Traditional methods, such as Kalman filters (KF) and Markov processes, rely heavily on past trajectories and hand-crafted features, which fail to account for the dynamic nature of construction sites. With the rising interest in data-driven approaches, several studies have explored different methods of trajectory prediction. Long Short-Term Memory (LSTM) network is one of the major methods used for forecasting future trajectories by leveraging both past individual and contextual information. However, one of the main limitations of LSTM is error accumulation, which limits the model from providing accurate results. Inspired by the success of the transformer model in natural language processing, this paper proposes the use of transformer encoder-decoder architecture with graph attention networks (GATs) to predict worker trajectories on construction sites. The temporal interactions of the workers are captured by the transformer model, while GAT captures the spatial relationships of the workers, which allow the model to build more comprehensive view of the workers behavior. The model is able to take 8 frames, covering 3.2 seconds, and predict the next 12 frames, covering 4.8 seconds, with an average displacement error (ADE) of 1.25 m and a final displacement error (FDE) of 2.3 m. The proposed model improves performance compared to traditional methods such as LSTM.