A context-augmented deep learning approach for worker trajectory prediction on unstructured and dynamic construction sites

Predicting workers' trajectories on unstructured and dynamic construction sites is critical to workplace safety yet remains challenging. Existing prediction methods mainly rely on entity movement information but have not fully exploited the contextual information. This study proposes a context-augmented Long Short-Term Memory (LSTM) method, which integrates both individual movement and workplace contextual information (i.e., movements of neighboring entities, working group information, and potential destination information) into an LSTM network with an encoder-decoder architecture, to predict a sequence of target positions from a sequence of observations. The proposed context-augmented method is validated using construction videos and the prediction accuracy achieved is 8.51 pixels in terms of final displacement error (FDE), with an observation time of 3 s and prediction time of 5 s-5.4% smaller than using the position-based method. Compared to conventional one-step-ahead predictions, the proposed sequence-to-sequence method predicts trajectories over multiple steps to avoid error accumulation and effectively reduces the FDE by 70%. In addition, qualitative analysis is conducted to provide insights to select appropriate prediction methods given different construction scenarios. It was found that the context-aware model leads to better performance comparing to the position-based method when workers are conducting collaborative activities.