Safe Navigation With Occluded Pedestrians: An Analytical Active Inference Approach

Abstract

Autonomous vehicles in urban environments must account for pedestrians potentially hidden behind static obstacles, where existence, position, and behavioral intent are all uncertain. We propose a fully analytical Active Inference framework that replaces particle-filter-based inference with a parametric Bernoulli-Gaussian Mixture belief. Belief updates are performed in closed form through Bayesian log-odds existence tracking and Gaussian truncation via bivariate normal CDFs. For planning, collision probability is computed as a continuously differentiable bivariate normal CDF integral conditioned on kinematic reachable sets, and an $\varepsilon$-contamination minimax formulation provides distributional robustness across pedestrian behavioral intents. Experiments across five adversarial pedestrian scenarios yield a 99.3% pass rate, a low collision speed of 0.56 m/s, and smoother control, outperforming both the particle-based predecessor and a state-of-the-art online POMDP solver, with a 2.4× planning speedup over the predecessor.