Logotyp: till Uppsala universitets webbplats

Pedestrian safety is a critical aspect of automated driving systems, making accurate pedestriantrajectory prediction an essential capability. In urban environments, pedestrian behavior isinfluenced by various contextual factors, including surrounding traffic, ego-vehicle motion, androad geometry. This thesis investigates whether incorporating such contextual information intodata-driven prediction models improves pedestrian trajectory prediction accuracy compared withapproaches that rely solely on the target pedestrian's motion history.To study this, several context-aware and context-agnostic model variants are evaluated usingmatched comparisons, ablations, targeted subsets, and crossing-intention prediction. Themodels range from simple LSTM-based predictors to complex Transformers. Pedestriansamples from nuScenes, Argoverse 2, and a Viscando intersection dataset are converted into ashared pedestrian-centered bird's-eye-view representation. The representation includes targetmotion, neighboring agents, ego-vehicle motion, and vectorized static scene context.The results show that contextual information can improve trajectory prediction, but the effect isstrongly dataset-dependent. Context provides consistent but moderate improvements onArgoverse 2, larger gains on the crossing-focused Viscando dataset, and only weakimprovements on nuScenes. Across experiments, target motion remains the dominant predictor,and more complex context-fusion architectures do not consistently outperform simpler models.However, crossing-intention prediction benefits clearly from context, showing that the scenerepresentation contains behaviorally meaningful information. Overall, the thesis builds aframework for incorporating context into pedestrian trajectory prediction and shows that contextcan support pedestrian prediction, but additional work is needed to make these gains consistentacross datasets and scenes. The implementation for this project is available in a public GitHubrepository: https://github.com/FloKnp/ContextAwarePedestrianTrajectoryPrediction.