MambaLSTM: A Spatio-Temporal Framework for Enhanced Traffic Accident Risk Prediction
The MambaLSTM framework predicts traffic accident risk by combining a squeeze-and-excitation temporal feature fusion module, a patch embedding module for adjacent spatial semantics, and a state-space Mamba block to capture global spatial correlations. The model also incorporates a specialized MambaLSTM unit to track both long- and short-term temporal dependencies across urban regions.
Why it matters — It prevents the introduction of noise during the fusion of spatial and temporal features while capturing global spatial relationships that traditional local-convolution models miss.