Risk‐Aware Multi‐Modal Pedestrian Trajectory Prediction With Scene‐Attentive Social Attention

ABSTRACT Pedestrian trajectory prediction is a fundamental task in intelligent transportation systems, where accurate forecasting must also remain safe and globally coherent in crowded multi‐agent scenes. However, low geometric error alone does not guarantee reliable predictions, since trajectories with small ADE/FDE may still involve unsafe interpersonal proximity or traverse spatially risky regions. To address this issue, this paper proposes SceneAware‐Att , a risk‐aware and scene‐attentive framework for multi‐agent trajectory prediction. The model jointly encodes historical motion and scene context, captures inter‐pedestrian dependencies through pedestrian‐wise self‐attention, and introduces a location‐conditioned risk‐aware attention mechanism to incorporate fine‐grained local spatial risk into trajectory decoding. To evaluate the method more comprehensively, we consider not only conventional accuracy metrics (ADE/FDE), but also safety‐oriented indicators, including collision rate (CR) and risk exposure (RE), together with joint multi‐agent metrics (JADE/JFDE) for scene‐level consistency assessment. Experiments on the ETH/UCY benchmark under the leave‐one‐out protocol show that the proposed method achieves the best overall performance among the compared methods: the single‐hypothesis deterministic setting, denoted as Ours (D), attains 0.19/0.30 ADE/FDE, while the best‐of‐ multi‐hypothesis setting, denoted as Ours (S), further improves to 0.17/0.28 with predicted hypotheses. Meanwhile, the proposed framework also reduces CR and RE and achieves better JADE/JFDE, indicating safer predictions and stronger global consistency of the predicted multi‐agent futures. Additional experiments on the Stanford Drone Dataset further demonstrate that the proposed design maintains favourable accuracy and safety performance under cross‐dataset evaluation. These results show that incorporating explicit spatial risk provides complementary benefits beyond optimising geometric error alone, supporting safer and more coherent pedestrian trajectory forecasting for safety‐critical ITS applications.