Limited Visibility and Uncertainty Aware Motion Planning for Automated Driving

Ömer Sahin Tas, Christoph Stiller

In the clean, simulated worlds of early autonomous vehicle research, sensors saw everything. In the real world, however, buildings block intersections, fog rolls in, and sensors have noise. If a self-driving car assumes the road is clear just because it can't see any obstacles, it’s a recipe for disaster.

This paper by Ömer Şahin Taş and Christoph Stiller from the FZI Research Center for Information Technology addresses how to plan motion when the world is occluded and uncertain.

TL;DR

The authors present a motion planning framework that proactively accounts for limited visibility and sensor uncertainty. By assuming a "worst-case" hidden object might exist just beyond the current line of sight, the planner ensures the vehicle always maintains a safe "fallback" option to stop, emulating cautious human driving behavior at blind intersections.

What Problem Does This Paper Address?

Most motion planners operate on a "perceived" environment. If the sensors don't detect an object, the planner assumes the space is empty. However, urban environments are full of occlusions—parked trucks, buildings, or sharp corners—that create blind spots.

The challenge is bridging the gap between an incomplete environment model and safe motion. A reactive system might slam on the brakes the moment a hidden car emerges from a blind spot, but a proactive system should slow down in anticipation of that possibility. The authors aim to formalize this "cautious" behavior into an optimization-based planner that handles both sensor noise (uncertainty) and physical obstructions (limited visibility).

Methods

The researchers utilize an optimization-based motion planner that minimizes a cost function (for comfort and progress) while adhering to safety constraints.

  1. Environment Modeling: They use a street-relative coordinate system and model the vehicle’s position and speed as Gaussian distributions to account for localization and perception noise.
  2. The "Virtual Object" Concept: For "Free Drive" scenarios with limited visibility, the planner treats the edge of the visible range as a potential location for a "virtual" stopped vehicle. The ego vehicle must maintain a speed that allows it to stop before hitting this hypothetical obstacle if it were to appear.
  3. Fallback Planning: A key concept is the "safe maneuver option." The planner ensures that for every planned trajectory, there exists a valid "fallback" (usually full braking) that remains collision-free even if the environment evolves in a worst-case manner during the system's processing delay ($t_d$).
  4. Intersection Logic:
    • Give-Way: If an intersection is occluded, the planner assumes a hypothetical vehicle is approaching at the speed limit from the hidden area.
  5. Right-of-Way: Even when the ego vehicle has the right of way, the planner checks for "uncompliant behavior" from others using the Intelligent Driver Model (IDM) to predict if another vehicle is failing to yield.
  6. Optimization: The constraints are solved using the Ceres solver, an open-source C++ library for modeling and solving large, complicated optimization problems.

Datasets

The paper does not use a standard public dataset like KITTI or Waymo. Instead, the approach was evaluated using a proprietary closed-loop simulation environment developed by the authors. This allowed them to test specific edge cases, such as varying sensor ranges and different levels of intersection occlusion.

Key Insights and Ideas

  • Proactive vs. Reactive: The "aha moment" is the mathematical formalization of caution. By treating visibility limits as hard constraints, the vehicle naturally "creeps" into blind intersections—not because it's programmed with a "creep" rule, but because the optimization requires a valid stop-path at all times.
  • The Surface-of-No-Return: The authors define a 3D space (path-time-speed) where certain states are "unsafe" because they leave the vehicle with no physical way to stop before a potential collision point. Safe planning means staying below this "surface."
  • Interaction Awareness: The system doesn't just look at distances; it uses the IDM to "guess" the intentions of other drivers. If a vehicle isn't slowing down as expected for a yield sign, the ego vehicle proactively prepares to brake, even if it has the legal right of way.

Key Results

  • Visibility-Speed Correlation: Simulations showed a direct relationship between sensor range and approach speed. With a 110m sensor range, the vehicle could maintain higher speeds through an intersection; at 30m, it significantly reduced speed to maintain its safety margins.
  • Smoothness: Despite the "worst-case" constraints, the use of a continuous optimization-based planner (rather than a discrete state machine) resulted in smooth velocity and acceleration profiles, avoiding jerky "emergency" maneuvers.
  • Safety Margins: The planner successfully maintained a collision probability below a defined threshold ($C$) by incorporating the standard deviation of sensor measurements into the distance constraints.

Limitations and Future Work

  • Longitudinal Focus: The current work focuses almost entirely on longitudinal control (braking and accelerating). It does not consider evasive steering maneuvers as a way to avoid collisions.
  • Map Dependency: The system relies on a highly accurate, up-to-date offline map (libLanelet). If the map is wrong about where an intersection is, the visibility logic might fail.
  • Computational Complexity: While Ceres is efficient, solving complex non-linear optimizations in real-time for every frame can be demanding.
  • Future Directions: The authors suggest extending the model to handle more complex maneuvers like lane changes and incorporating more diverse road users like cyclists and pedestrians, who have different dynamic constraints than cars.

Summary updated

Abstract

Adverse weather conditions and occlusions in urban environments result in impaired perception. The un-certainties are handled in different modules of an automated vehicle, ranging from sensor level over situation prediction until motion planning. This paper focuses on motion planning given an uncertain environment model with occlusions. We present a method to remain collision free for the worst-case evolution of the given scene. We define criteria that measure the available margins to a collision while considering visibility and interactions and consequently integrate conditions that apply these criteria into an optimization-based motion planner. We show the generality of our method by validating it in several distinct urban scenarios.
BibTeX
@inproceedings{tas2018limited,
  author = {Ömer Sahin Tas and Christoph Stiller},
  title = {{Limited Visibility and Uncertainty Aware Motion Planning for Automated Driving}},
  booktitle = {Intelligent Vehicles Symposium},
  year = {2018},
  url = {https://openreview.net/forum?id=cz5BcWJiwp},
}

Loading related content...

Figure