Paying Attention to Infrastructure in Trajectory Predictions : Learning Movement Through Industrial Spaces

Sammanfattning

This thesis investigates single-agent trajectory prediction in industrial environments using a transformer-based model with explicit environmental cross-attention, named Environment-Aware Transformer (EAT). The environment is represented as a 2D semantic grid enriched with derived geometric and traversability-related feature channels. The aim is to examine whether explicit environmental conditioning improves trajectory prediction and whether such conditioning supports more physically plausible trajectory predictions in structured industrial settings. The model is evaluated in two experimental settings. The first examines in-distribution prediction performance through comparison with a separately trained non-environment-aware transformer (NEAT) baseline and a constant-velocity Kalman filter. The second evaluates generalization to a previously unseen industrial environment. Performance is assessed using Average Displacement Error (ADE), Final Displacement Error (FDE), Trajectory Collision Rate (TCR), qualitative trajectory examples, and attention-based analysis. The results show that environmental conditioning does not provide a consistent overall improvement in ADE and FDE compared with NEAT. However, EAT performs better on curved trajectories in the unseen environment and produces lower TCR across the evaluated settings. These findings imply that environmental context may be most useful when motion is shaped by surrounding infrastructure and when physical plausibility matters. The thesis therefore highlights the limitations of relying only on pointwise displacement errors and provides an early step toward using trajectory prediction to evaluate industrial layout changes virtually before they are implemented.