<![CDATA[This position paper argues that current Vehicle-to-Everything (V2X) safety strategies have reached a plateau due to their reliance on rigid, single-prediction models that fail to account for the heterogeneous movement characteristics of road users at intersections. A position paper, as recognized in the scientific literature, advocates for a specific research direction through systematic evidence synthesis and preliminary validation, rather than presenting exhaustive experimental results. Our systematic review of 24 studies spanning 2020–2025 reveals a critical tension: while Long Short-Term Memory (LSTM) models dominate the research landscape with a 42% share, they are computationally expensive for edge deployment, whereas lighter kinematic models exhibit prediction errors exceeding 3.60 meters in complex settings. To resolve this tension, we position a hybrid Kalman Filter (KF) and LSTM architecture implemented within the Sensing, Connectivity, Intelligence, and Actuating (SCIA). The KF, augmented with a formally specified stop-aware kinematic logic, handles structured vehicle trajectories, while LSTM exclusively handles the stochastic behavior of pedestrians. Preliminary co-simulation results (SUMO, OMNeT++, Artery; 500-second scenario; 14,000+ samples per horizon) demonstrate that the hybrid model reduces pedestrian trajectory RMSE from 5.07 m to 1.23 m (75.74% reduction) and vehicle RMSE from 23.69 m to 21.57 m (8.94% reduction) at the 5-second horizon compared to a linear regression baseline. These results confirm that asymmetric model delegation is architecturally superior to uniform single-model approaches and validate the proposed direction for future V2X safety application development. Full collision detection performance evaluation constitutes the primary contribution of a companion experimental study currently in preparation.]]>
