Pedestrian–vehicle interactions in urban traffic are difficult to model because they unfold over short time scales, depend on local geometric context, and involve ambiguity between observed motion and latent intention. This thesis investigated whether short-horizon pedestrian crossing-corridor entry could be predicted from real-world trajectory data and how the conclusions depended on key design choices, especially the future horizon. Using stationary trajectory data from Viscando collected in an urban setting in Gothenburg, Sweden, the study defined an operational target as whether a pedestrian entered a predefined crossing corridor within a fixed future horizon. Horizons of 2, 3, and 5 seconds were compared, with 3 seconds selected as the main reporting horizon.
The approach combined interpretable feature engineering, logistic-regression baselines with staged interaction-feature ablation, a feed-forward MLP, and a compact GRU sequence-model comparison. Transparency was pursued both through interpretable model design and through post-hoc permutation-importance analysis applied to the strongest non-linear model. The intended application is a decision-support framework for traffic engineers, road-safety analysts, or ADAS developers requiring both accurate predictions and understandable model behavior.
Pedestrian-only motion and corridor-relative geometry provided a strong baseline. On the matched vehicle-context subset, richer interaction structure, especially relative vehicle geometry, was more informative than simple proximity. The interaction-aware MLP was the strongest overall point-estimate result, though the within-family pedestrian-only versus interaction-aware difference was not statistically confirmed. The compact GRU sequence branch did not outperform the engineered tabular MLP under the present representation and data conditions. Horizon choice materially affected class balance, performance, and which interaction signal was most useful, shifting from vehicle motion at 2 seconds, to relative geometry at 3 seconds, to local density at 5 seconds. Taken together, the findings show not only that short-horizon crossing-corridor entry can be predicted from stationary trajectory data, but also that target formulation, horizon choice, and feature representation materially shape what can be learned from the data.
2026. , s. 61