Bespoke Deep Learning Approaches for Pedestrian Trajectory Prediction

Human trajectory prediction involves forecasting future movements based on past positions and is crucial in various fields, including socially-aware robots, intelligent tracking systems, and autonomous vehicles. This work aims to learn representations from observed scenes and apply knowledge distillation to transfer insights from a teacher network to a student network. To achieve this, we integrate social and semantic maps, along with goal/waypoint heatmaps, into a multi-modal temporal backbone. Knowledge is transferred from a teacher network, which predicts short-termtrajectories based on long- term observations, to a student network, which predicts long-term trajectories from short-term data. We conducted extensive experiments with different lengths of teacher inputs and training data over longer time horizons, demonstrating that our model outperforms the state-of-the-art on the SDD and inD datasets in both short-term and long-term scenarios.