Source-linked AI summary
Occlusion-Aware Risk Assessment for Autonomous Driving in Urban Environments
Ming-Yuan Yu, Ram Vasudevan, Matthew Johnson-Roberson
TL;DR
Urban sensors leave regions unobserved because of limited range and occlusion, while human drivers anticipate unseen traffic. The paper quantifies risk in those regions using known road layouts; in evaluation, collision rates fell by up to 4.8× and comfort improved by up to 10×.
Problem
Limited sensor range and occlusion leave urban vehicles with unseen regions, and drivers anticipate possible traffic there when deciding whether to turn.
Source: §I · p. 1Method
The algorithm performs probabilistic risk assessment of observed and unobserved regions at urban intersections and can integrate with different planning approaches.
Source: §II · p. 2, B. Planning · p. 4Results
Collision rate decreases by up to 4.8× and comfort increases by up to 10× compared with the baseline.
Source: §VI · p. 6, A. Baseline 1 - Collision Rate · p. 6, B. Baseline 1 - Ride Comfort · p. 6Takeaways & Limitations
The assessed risk can inform vehicle planning, and intersection collision rates can help a high-level planner avoid dangerous intersections.
Source: §II · p. 2, D. Metrics · p. 5
Abstract
from arXiv · showhide
Navigating safely in urban environments remains a challenging problem for autonomous vehicles. Occlusion and limited sensor range can pose significant challenges to safely navigate among pedestrians and other vehicles in the environment. Enabling vehicles to quantify the risk posed by unseen regions allows them to anticipate future possibilities, resulting in increased safety and ride comfort. This paper proposes an algorithm that takes advantage of the known road layouts to forecast, quantify, and aggregate risk associated with occlusions and limited sensor range. This allows us to make predictions of risk induced by unobserved vehicles even in heavily occluded urban environments. The risk can then be used either by a low-level planning algorithm to generate better trajectories, or by a high-level one to plan a better route. The proposed algorithm is evaluated on intersection layouts from real-world map data with up to five other vehicles in the scene, and verified to reduce collision rates by 4.8x comparing to a baseline method while improving driving comfort.
I. INTRODUCTION
Finite sensor range and opaque obstacles leave urban vehicles with unobserved regions. The paper proposes quantifying risk from limited sensing and geometric occlusion, reflecting how human drivers anticipate unseen traffic and supporting safer, more comfortable driving.
I. INTRODUCTION
- Finite range, sensitivity, and angular resolution limit object detection, while opaque objects create large regions sensors cannot see.The paper illustrates this setting at an intersection.Source: §I · p. 1
- Drivers use a mental map of unseen space and wait for a sufficient traffic gap before turning.They may also pull forward to improve visibility.Source: §I · p. 1
- The proposed algorithm quantifies risk from limited sensing and geometric occlusion to encode this form of human driving.It is intended to support safe urban driving with improved ride comfort and is agnostic to how vehicles make decisions.Source: §I · pp. 1–2
II. RELATED WORK
Prior work either tracks already-detected occluded agents or faces tractability and evaluation limits; this paper instead estimates probabilistic risk at urban intersections and keeps that assessment independent of the planning algorithm.
II. RELATED WORK
- Occlusion-tracking methods require an initial detection and do not explicitly assess traffic that is hidden or outside sensor range and never detected.The cited approaches maintain tracks for targets entering occluded regions, leaving potential incoming traffic unmodeled.Source: §II · p. 2
- POMDP approaches can implicitly handle probabilistic occlusion, but state-space reduction limits their applicability, particularly for real-time high-speed driving.This paper instead quantifies environmental risk rather than risk associated with ego-vehicle actions, and remains compatible with POMDP planning.Source: §II · p. 2
- The closest prior methods either over-approximate incoming traffic non-probabilistically or show limited results with one additional vehicle, unlike this work’s crowded real-map intersections.The paper evaluates up to five other vehicles in the scene.Source: §II · p. 2
III. METHOD
- The paper assesses probabilistic risk for observed and unobserved regions and makes the resulting Cartesian-space risk usable by planning algorithms.The risk can serve as a primary or auxiliary cost; the method also demonstrates integration with a simple optimization-based planner.Source: §II · p. 2, §III · p. 2, A. Risk Assessment Over Cartesian Space · p. 3
A. Risk Assessment Over Cartesian Space
- Known maps and ego location define an observable polygon, but current free-space estimates alone provide little information about long-term risk from vehicles that may emerge unseen.The method uses particles to represent potential vehicle locations originating in unobserved regions.Source: A. Risk Assessment Over Cartesian Space · pp. 2–3
- For each lane spline, the method samples particles in unobserved segments using uniform position and speed distributions, assuming no prior information about either.The speed range is bounded by the minimum and maximum speeds of other vehicles; a more descriptive distribution is possible when prior information exists.Source: A. Risk Assessment Over Cartesian Space · p. 3
- Particles are propagated at constant speed for a forecast horizon, then laterally offset to represent vehicle size and within-lane displacement.More sophisticated variable-speed models are possible at greater computational cost; uniform lateral offsets reflect no prior position information.Source: A. Risk Assessment Over Cartesian Space · p. 3
- Observed vehicles are modeled similarly, and the scene’s risk distribution unions the particle sets for use in planning costs.Treating observed vehicles this way is conservative when their intent and speed are unknown or noisy.Source: A. Risk Assessment Over Cartesian Space · pp. 3–4
B. Planning
The planner incorporates occlusion-aware risk into an optimization-based approach that balances safety with desired speed. It selects acceleration by minimizing a weighted objective, with smaller λ favoring more conservative behavior.
B. Planning
- The planner integrates the risk assessment into an optimization-based planner, an approach the authors say can be used with any optimization-based planner.The rudimentary planner can be replaced by another cost-based planner because the technique is agnostic to the planning approach.Source: B. Planning · p. 4
- The safety cost J1(aego) uses particle potential values that are positive when a predicted particle is within the ego lane.The planner also considers a speed cost J2(aego) to drive the ego vehicle toward its desired speed.Source: B. Planning · p. 4
- The planner finds the optimal acceleration over the interval from t to t+Tp by solving an optimization problem, with smaller λ favoring more conservative behavior.The formulation includes limits on ego speed and acceleration; λ balances the planner's behavior, and the authors report the Table I value as a compromise between safety and efficiency.Source: B. Planning · p. 4
IV. EVALUATION
The evaluation compares the proposed method with a baseline that models only observed vehicles, focusing on difficult maneuvers such as unprotected left turns.
IV. EVALUATION
- The evaluation compares the proposed method with a baseline modeling only observed vehicles, focusing on difficult maneuvers such as unprotected left turns.Source: §IV · p. 4
A. Simulation
The simulations use random vehicle scenarios at four-way, un-signaled intersections, including layouts from real-world maps. The setup adds buildings to create heavy occlusion and tests unprotected left turns.
A. Simulation
- Each simulated scene includes five other vehicles traveling random routes at constant speeds from 4 to 12 m/s.Valid trajectory combinations are generated by rejection sampling to avoid collisions or overlap among the simulated vehicles.Source: A. Simulation · p. 4
- The evaluation focuses on four-way, un-signaled intersections, using synthetic layouts or geometry from 73 real-world intersections around Ann Arbor, Michigan.The real-world intersections were extracted from OpenStreetMaps; the approach is described as conceptually generalizing to T- or Y-junctions.Source: A. Simulation · pp. 4–5
- Fig. 4 illustrates collision rates overlaid on a map of 12 intersections, with a high-level planner taking the route with lower collision rates.The illustrated route is (2, 0) → (2, 1) → (1, 1) → (0, 1) → (0, 2) → (0, 3).Source: Figure 4 · p. 5
- Buildings are placed 2 m beyond the driving-surface boundary to simulate heavy occlusion, and the ego vehicle starts 15 m before the stopline at 10 m/s for an unprotected left turn.Particle density is held at 215 particles per 100 m as unobserved-segment lengths vary.Source: A. Simulation · p. 5
B. Baseline 1
The baseline predicts risk using the same method and planning algorithm as the proposed approach, but only for regions intersecting observed vehicles.
B. Baseline 1
- The baseline predicts risk only for regions intersecting observed vehicles, while using the same risk-prediction method and planning algorithm as the proposed approach.Source: B. Baseline 1 · p. 5
C. Baseline 2
The authors describe [6] as the only method they know of that combines risk assessment and planning, but note that it showed only a few examples. They recreate its scenarios to compare speed and acceleration profiles.
C. Baseline 2
- The scenarios from [6] are recreated to compare speed and acceleration profiles.The comparison follows [6]’s scenarios rather than relying only on the few examples in that paper.Source: C. Baseline 2 · p. 5
- The authors characterize [6] as the only method they know of that describes both risk assessment and planning in one article.They note that [6] presents only a few examples.Source: C. Baseline 2 · p. 5
- The recreated scenarios use the same map, initial speed, desired speed, and acceleration limits as [6].The other vehicle’s initial position and speed are estimated from plots in [6].Source: C. Baseline 2 · p. 5
D. Metrics
The evaluation compares collision risk under matched simulations and assesses discomfort from vehicle speed and acceleration profiles. The section also describes how intersection-level collision rates can inform route planning.
D. Metrics
- Baseline 1 runs 2,000 random unprotected-left-turn scenarios per intersection, then repeats the same experiments with identical trajectories using the proposed method.The comparison is made against the baseline method.Source: D. Metrics · p. 5
- Among all 73 real-world intersections, Fig. 5’s CDFs show our method outperforms the baseline: the 95th-percentile collision rate falls from 12.64% to 2.61%, and discomfort from 0.4925 to 0.1043.Source: Figure 5 · p. 5
- Intersection-level meta-collision rates can inform a high-level planner choosing a route across a city.Overlaying the rates on their intersections lets the planner avoid dangerous intersections.Source: D. Metrics · p. 5
- Discomfort is quantified from speed and acceleration profiles using a continuous discomfort score.The score is introduced because the authors report no common computational metric for ride comfort; prior work typically reports acceleration and jerk thresholds [18].Source: D. Metrics · p. 5
- The discomfort threshold is set to half the maximum deceleration, and the evaluated scenarios include an empty intersection and one with a vehicle approaching from the left.The threshold is positive, and the scenarios are studied under occlusion.Source: D. Metrics · p. 5
V. RESULTS
The results section characterizes collision-rate and discomfort-score distributions using medians and 95th percentiles, treating the latter as near-worst cases.
V. RESULTS
- The results report the median and 95th percentile of both collision-rate and discomfort-score distributions.The distributions are shown as CDFs in Figure 5a and 5b.Source: §V · p. 5
- The median collision rate and discomfort score represent nominal behavior.Source: §V · p. 5
- The 95th percentile of each distribution represents near-worst-case behavior.Source: §V · p. 5
A. Baseline 1 - Collision Rate
The proposed method reduces collision rates across synthetic and real-world intersections. The real-world intersection collision-rate distribution can also be overlaid on a map for route-planning use.
A. Baseline 1 - Collision Rate
- Collision rates for a subset of intersections are overlaid on a real-world map, where a high-level planner can avoid dangerous intersections such as the two yellow ones.Source: Figure 6 · p. 6
- 4.8×: across 73 real-world intersections, the 95th-percentile collision rate decreases from 12.64% to 2.61%.The median collision rate decreases 3.7×, from 5.40% to 1.45%.Source: A. Baseline 1 - Collision Rate · p. 6
- The distribution of simulated collision rates across the 73 real-world intersections can provide route-planning information and support safety assessment before a journey.The passage also identifies a systematic, quantitative use for urban planners and civil engineers assessing intersection safety for autonomous vehicles.Source: A. Baseline 1 - Collision Rate · p. 6
B. Baseline 1 - Ride Comfort
The proposed method improves collision rate and ride comfort in the reported evaluations. Its anticipatory slowing also produces more consistent speed and acceleration profiles than the baselines.
B. Baseline 1 - Ride Comfort
- 2.9× (0.0795 to 0.0271) median and 10× (0.4687 to 0.0466) 95th-percentile discomfort-score reductions occurred at the synthetic intersection.Across real-world intersections, median discomfort score fell 3×, from 0.0849 to 0.0284, and the 95th percentile fell 4.7×, from 0.4925 to 0.1043. The baseline shows abrupt braking and deceleration down to −8 m/s2, while the proposed method predominantly stays above −4 m/s2 with smaller variations.Source: B. Baseline 1 - Ride Comfort · p. 6, Figure 8 · p. 7
C. Baseline 2
- Both [6] and the proposed method traverse the intersection without collision, but the proposed method accelerates and decelerates gracefully while [6] shows large jerk.In the first scenario, the proposed method obtains zero acceleration while the baseline unnecessarily decelerates.Source: C. Baseline 2 · p. 6
VI. CONCLUSIONS
- Up to 4.8× collision-rate reduction and up to 10× comfort improvement are reported across one synthetic and 73 real-world intersections.The evaluation included up to five other vehicles in the scene.Source: §VI · p. 6
- Future work includes incorporating other vehicles' information for cooperative planning and handling heterogeneous traffic with vehicles and pedestrians.Source: §VI · p. 6