Geometry-informed Bayesian deep learning for proactive driving risk assessment: a full spatiotemporal probabilistic...

Proactive driving risk assessment is essential for autonomous vehicles and advanced driver-assistance systems. Traditional conflict indicators and risk field approaches are limited in capturing inherent motion uncertainties and multi-dimensional crash risks in a unified manner. Although deep learning facilitates probabilistic trajectory prediction, purely data-driven models often lack geometric plausibility and neglect epistemic uncertainty, yielding implausible predictions and unreliable risk quantification. To address these limitations, a unified proactive driving risk assessment framework is proposed by integrating a geometric-informed Bayesian heteroscedastic convolutional social long short-term memory (GIBH-CS-LSTM) network with a full spatiotemporal probabilistic risk field (FSPRF) approach. Within the predictive architecture, aleatoric and epistemic uncertainties are simultaneously quantified through heteroscedastic outputs and variational inference. Geometric boundary constraints and an adaptive dual-entropy regularization (ADER) mechanism are incorporated to ensure spatial reliability and maintain the anisotropic representation of motion variances. Based on these predictive distributions, the FSPRF methodology unifies the quantification of synchronous interaction risks, asynchronous conflict risks, and single-vehicle run-off-road risks. Specifically, based on the risk field intensity of vehicles, a set-based instantaneous collision risk index (ICRI) is defined to evaluate the interaction intensity between extended rigid bodies. This metric employs a whitening transformation to convert the minimum Mahalanobis distance between sets into a minimum geometric distance calculation within an isotropic whitened space, rectifying risk underestimation near vehicle boundaries. The framework is validated utilizing vehicle trajectory data extracted from freeway curved and straight sections. Quantitative evaluations demonstrate that the GIBH-CS-LSTM model mitigates road departures and improves trajectory accuracy. The unified FSPRF captures lateral and longitudinal crash risks and enables direct aggregation of risk exposure, providing a more comprehensive multidimensional evaluation compared to traditional conflict indicators. Ultimately, this framework provides a quantitative foundation for developing proactive collision avoidance strategies.