Occlusion-Aware Risk Assessment for Autonomous Driving in Urban Environments

Yu, Ming-Yuan; Vasudevan, Ram; Matthew Johnson‐Roberson · 2019 · OpenAlex-citations

DOI: 10.1109/lra.2019.2900453

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of safe autonomous navigation in urban environments where sensor limitations and physical occlusions create unobserved regions. The authors argue that relying solely on visible data is insufficient for safety, as vehicles may emerge from hidden areas. To mitigate this, the paper proposes an occlusion-aware risk assessment algorithm that quantifies the probability of collision with unseen vehicles by leveraging known road layouts and geometric constraints. This approach allows autonomous systems to anticipate future risks and adjust their behavior proactively, improving both safety and ride comfort. The method utilizes High Definition (HD) maps to identify road geometry and define an observable polygon based on sensor range and field of view. For unobserved road segments, the algorithm employs a particle filter approach to forecast potential vehicle locations. Particles representing potential vehicles are sampled from uniform distributions along road centerlines, assuming constant speeds within known nominal ranges. These particles are propagated forward in time, and lateral offsets are added to account for vehicle width and lane positioning. This generates a probabilistic distribution of risk over the Cartesian space, covering both observed and unobserved areas. This risk map is then integrated into an optimization-based planning algorithm as a cost function, encouraging the ego vehicle to decelerate or adjust its trajectory to avoid high-risk zones. The algorithm was evaluated through simulations on both synthetic and 73 real-world intersection layouts derived from OpenStreetMap data around Ann Arbor, Michigan. The ego vehicle performed unprotected left turns in scenarios with up to five other vehicles. The proposed method was compared against a baseline that only assessed risk from observed vehicles and a second baseline from prior literature. Results demonstrated significant improvements in safety metrics. At synthetic intersections, the collision rate decreased by 4.1 times (from 5.75% to 1.40%). Across the real-world intersections, the median collision rate dropped by 3.7 times, and the 95th percentile of collision rates decreased by 4.8 times, falling from 12.64% to 2.61%. Additionally, the method improved ride comfort by reducing unnecessary deceleration, as evidenced by lower discomfort scores compared to the baselines. The significance of this work lies in its ability to provide a control-agnostic risk assessment tool that enhances autonomous driving safety in complex, occluded environments. By explicitly modeling the risk of unseen traffic, the algorithm enables vehicles to navigate intersections more cautiously and comfortably. The findings suggest that incorporating geometric priors and probabilistic forecasting of occluded regions can substantially reduce collision probabilities, offering a robust solution for urban autonomous driving scenarios where sensor visibility is limited.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness