Spatio-Temporal Feature Learning Fusion and Visual Scene Endpoint Prediction for Pedestrian Trajectory Prediction

Pedestrian trajectory prediction plays a crucial role in computer vision with a broad range of applications, which is a challenging task due to temporal dependence of pedestrians and their spatial interactions, especially in multi-pedestrian scenarios where temporal and spatial information are intricately coupled. Addressing how to enable models to learn complex spatio-temporal correlations is a key challenge. Moreover, beyond historical movement trajectories, predicting pedestrian trajectories requires understanding environmental contexts and potential endpoints. To address these challenges, we propose a novel pedestrian trajectory prediction model named Spatio-Temporal Feature Learning Fusion and Visual Scene Endpoint Prediction(STEP). The model comprises two modules: spatio-temporal feature learning module and visual scene endpoint prediction module. In the spatio-temporal feature learning fusion module, two branches are employed to learn temporal and spatial features separately, followed by adaptive fusion through a gated network. To effectively simulate pedestrian-environment interactions, the visual scene endpoint prediction module employs background image semantic segmentation to predict scene-consistent endpoints. We conducted extensive experiments on two public datasets, ETH-UCY and SDD, and the results demonstrate the proposed approach’s effectiveness.