Pedestrian crossing intention prediction using motion and interaction features : For advanced driver assistance systems and autonomous driving applications

Sammanfattning

Pedestrian safety remains one of the most critical challenges in modern transportation systems, particularly with the rapid development of autonomous driving technologies and intelligent traffic environments. Predicting whether a pedestrian will cross the road is a safety- critical task for advanced driver assistance systems and autonomous vehicles. The ability to accurately predict pedestrian crossing intention before the actual crossing occurs can significantly improve vehicle response time and reduce potential accidents. In this project, we investigate early pedestrian crossing-intention prediction using real-world trajectory data from a Swedish urban intersection. In this thesis, crossing intention is operationalized as entry into a predefined crossing zone for at least two consecutive frames, corresponding to a 0.5-second entry rule. We compare sequence models (GRU, LSTM), an attention-based model (Transformer), and an interaction-aware model (GNN), with a Kalman filter as a trajectory baseline. Multiple experiments were conducted to evaluate model performance and investigate how early crossing intentions can be reliably predicted before the actual crossing event. We explicitly evaluate motion-only features against interaction-augmented motion features to quantify the effect of pedestrian-vehicle context. Furthermore, the thesis evaluates prediction performance at different time horizons (early-horizon protocol at 0.5 s, 1.0 s, 1.5 s, and 2.0 s) prior to the crossing event in order to determine how early reliable predictions can be made. Motion-only trajectory features are found to be highly predictive of crossing intention across all tested models. The trajectory-first feature-extraction protocol provides consistent improvements over baseline classifiers, while tabular interaction features did not produce reliable gains in this dataset. The GNN variants were evaluated at a single fixed setting only (T_obs = 2.0 s, H = 1.0 s) and their results are therefore not directly comparable to the 16-setting averages reported for the recurrent and attention-based classifiers. Horizon-dependent degradation was not observed for intention classification, though trajectory prediction error increased consistently with longer prediction horizons.