Pedestrian trajectory prediction based on spatio-temporal attention mechanism

Mining and modeling city residents’ long-term and short-term preferences from their historical trajectory data is a key issue in trajectory prediction. People show different preferences in the long and short terms, but existing methods fail to fully consider the periodicity of residents’ long-term preferences and the abruptness of their short-term preferences, leading to unsatisfactory prediction results. To tackle this issue, we proposes a novel Spatio-Temporal Attention Mechanism Model to effectively capture the long- and short-term preferences of residents. Specifically, for the long-term preferences, we integrate the temporal context information generated during the user’s movement through an improved time-weighting operation to obtain the representation of temporal preferences. Then, we utilize a self-attention mechanism that combines geographic location factor to obtain the representation of long-term preferences. And for the short-term preferences, the dynamic programming is employed to capture the abruptness of residents’ access and obtain the representation of short-term preferences. Our experimental results on three benchmark datasets show the superiority of our proposed model over the state-of-the-art trajectory prediction models. It outperforms them in terms of Recall and NDCG, while also demonstrating superior robustness.