Time-Aware Goal-Conditioned Pedestrian Trajectory Prediction

Accurate pedestrian trajectory prediction is essential for ensuring safe and anticipatory behavior in autonomous systems. We introduce a time-aware goal-conditioned pedestrian trajectory prediction framework aimed at improving behavior modeling in autonomous vehicle (AV) driving scenarios. Unlike traditional goal-conditioned approaches that rely solely on spatial features, our method incorporates time-of-day as an additional contextual signal to better reflect real-world pedestrian intent. Our architecture consists of two LSTM-based modules: a goal estimator that infers destination probabilities based on past trajectory and time, and a trajectory predictor that generates future positions conditioned on the soft goal distribution. We validate our framework using synthetic data with time-varying goal preferences, simulating temporally dependent pedestrian behavior. The results demonstrate that incorporating temporal features enhances prediction performance, particularly in scenarios with ambiguous or overlapping spatial cues. This proof-of-concept highlights the potential of using temporal context to improve pedestrian modeling, especially in infrastructure-supported environments such as smart intersections or testbeds. ... mehrOur approach lays the groundwork for more context-aware and realistic evaluation pipelines in AV research. While our evaluation is currently limited to synthetic data, we discuss implications and outline the requirements for real-world validation.