Probabilistic Models for Predicting Hidden Road Users

/ General / By

Occlusions create uncertainty about whether and where an unseen road user might appear. Probabilistic world models turn that uncertainty into actionable information by estimating a spatial probability distribution over the hidden space and propagating it forward in time. Below are practical approaches commonly used in autonomous driving.

Occupancy grids and predicted-occupancy grids

An occupancy grid discretizes the scene into cells and maintains for each cell a probability that it is occupied. A predicted-occupancy grid (POG) extends this idea by forecasting those probabilities over a short time horizon, incorporating motion models and map priors (e.g., sidewalks, lanes). POGs give planners a spatial risk map: high-probability cells near a crosswalk indicate places to slow or stop.

Bayesian filters and belief propagation

Bayesian filtering (binary Bayes filter, Bayesian Occupancy Filter) updates cell occupancy beliefs when sensors observe free or occupied evidence and when beliefs are propagated by a motion model. These filters naturally fuse multiple sensors and can incorporate prior information such as road semantics or typical pedestrian emergence points.

Particle filters and multi-hypothesis sampling

When dynamics or environment complexity make closed-form updates impractical, Sequential Monte Carlo (particle) filters represent hidden objects or dynamic cell states with samples. Each particle encodes a plausible state (position, velocity, intent); resampling and weighting concentrate probability on likely hypotheses. Particle-based predicted-occupancy maps capture multimodal possibilities (e.g., pedestrian might step onto crosswalk or remain on sidewalk) that Gaussian approximations would miss.

Random finite set (RFS) and PHD approaches

RFS-based filters (e.g., Probability Hypothesis Density, PHD) treat the number and states of hidden actors as a random set rather than separate tracked objects. This is useful when occluded regions may contain zero, one, or multiple small actors and when explicit data association is unreliable.

Combining map priors and scene context

All probabilistic models improve realism by conditioning on map semantics and traffic rules: sidewalks, crosswalks, building footprints, and lane geometry constrain feasible emergence locations and motion directions. Models also use contextual cues—nearby braking vehicles, oriented bicycles, parked vans—to reweight hypotheses about hidden actors.

Output for planning: risk metrics and conservative constraints

Perception modules export risk as cell occupancy probabilities, time-indexed collision likelihoods, or “criticality” scores. Planners convert those into conservative behaviors: reduced speed, stopping before the occlusion, viewpoint maneuvers (creeping forward to reduce occlusion), or larger safety buffers. The key is calibrated probabilities so planners neither overreact nor under-protect.

Computational trade-offs and real-time considerations

High-resolution POGs and dense particle sets improve fidelity but increase CPU/GPU cost. Practical systems use hybrid strategies: coarse-grid probabilistic maps for distant/occluded regions, object-level trackers for visible actors, and selective particle sampling near critical zones (crosswalks, intersections).

Validation and safety assurance

Quantitative validation compares predicted occupancy against ground truth from instrumented scenes or recorded intersections, measuring calibration (predicted probability vs. observed frequency) and worst-case miss rates. Safety cases typically require conservative thresholds and explicit handling of low-probability, high-consequence hypotheses (e.g., child running into street).

In short, probabilistic occupancy representations—implemented with Bayesian filters, particle methods, or RFS/PHD variants and informed by map priors and scene cues—are the practical tools used to quantify and act on the “large uncertainty about what is hidden.” They turn invisible risk into spatially and temporally explicit inputs for safe planning.

Sources

Prejemajte vse najnovejše novice in informacije v svoj nabiralnik.

l Slovenščina