Fluid-inspired field representation for risk assessment in road scenes

Xuanpeng Li, Lifeng Zhu, Qifan Xue, Dong Wang, Yongjie ZhangComputational 2020 · DOI

Abstract

Prediction of the likely evolution of traffic scenes is a challenging task because of high uncertainties from sensing technology and the dynamic environment. It leads to failure of motion planning for intelligent agents like autonomous vehicles. In this paper, we propose a fluid-inspired model to estimate collision risk in road scenes. Multi-object states are detected and tracked, and then a stable fluid model is adopted to construct the risk field. Objects’ state spaces are used as the boundary conditions in the simulation of advection and diffusion processes. We have evaluated our approach on the public KITTI dataset; our model can provide predictions in the cases of misdetection and tracking error caused by occlusion. It proves a promising approach for collision risk assessment in road scenes.

Introduction

Autonomous vehicles (AVs) operate in highly dynamic environments where sensor data is often noisy or incomplete. A significant challenge in motion planning is predicting how a scene will evolve, especially when objects are temporarily occluded—such as a car passing behind a bus. Traditional risk assessment methods, like Time-to-Collision (TTC), often struggle with non-linear trajectories or complex road geometries. Conversely, high-fidelity trajectory prediction models can be computationally expensive and prone to failure if an object is momentarily lost by the tracker.

This research addresses these issues by proposing a "fluid-inspired" risk representation. By treating collision risk as a continuous field that flows and spreads like smoke, the model provides a computationally efficient and robust framework for real-time risk assessment.

Methodology

The proposed pipeline moves from discrete 3D object detection to a continuous, dynamic risk field using principles from fluid mechanics.

1. Detection and Tracking

The system employs PointRCNN to detect 3D bounding boxes from LiDAR point clouds. These detections are then processed by a 3D Kalman Filter to track the position and velocity of each object over time, providing the necessary inputs for the fluid simulation.

2. The Fluid-Inspired Risk Field

The core of the approach is the application of the Navier-Stokes equations (specifically the "Stable Fluids" method) to a 2D grid representing the bird’s-eye view of the road. Each tracked vehicle acts as a source of "virtual smoke" or risk density. The field evolves through several physical processes:

  • Advection: The risk density is transported according to the vehicle's velocity vector.
  • Diffusion: The risk spreads out over time, naturally representing the increasing spatial uncertainty of the object's future position.
  • Anisotropy: To better reflect driving dynamics, diffusion is made anisotropic, meaning the risk spreads more significantly in the direction of travel (forming a teardrop shape) than laterally.
  • Dissipation: A decay factor is applied to ensure that older, less certain information gradually fades from the map.

Results

The model was evaluated using the KITTI tracking dataset, a benchmark for real-world urban and highway driving.

Robustness to Occlusion

The most significant advantage of the fluid model is its inherent "memory." In scenarios where a vehicle is lost by the tracker (e.g., due to sensor occlusion), the risk field associated with that vehicle does not immediately disappear. Instead, it continues to advect and diffuse based on its last known state. This allows the AV to maintain a high-risk zone where the hidden vehicle is likely to be, significantly improving safety over traditional occupancy grids that might treat the area as "clear."

Real-Time Performance

Despite the underlying complexity of partial differential equations, the "Stable Fluids" implementation is highly efficient. The system can update a 512x512 resolution risk map in approximately 25ms on a single CPU thread, achieving a frame rate of 40 FPS. This performance is well-suited for integration into real-time autonomous driving software stacks.

Discussion and Future Work

The fluid-inspired model offers a physically grounded prior for uncertainty that aligns well with human intuition. However, several avenues for future research remain:

  • Object Diversity: Expanding the model to include pedestrians and cyclists, who have distinct movement patterns and risk profiles.
  • Static Constraints: Integrating road geometry (e.g., lane markings, curbs) as boundary conditions in the fluid simulation to prevent risk from "bleeding" into impossible locations like sidewalks.
  • Environmental Adaptability: Investigating how extreme weather conditions might affect fluid parameters like dissipation and diffusion rates.

In conclusion, the fluid-inspired field representation provides a robust and efficient solution for collision risk assessment, particularly in the face of the sensing uncertainties inherent in real-world road scenes.

Summary updated

Abstract

Prediction of the likely evolution of traffic scenes is a challenging task because of high uncertainties from sensing technology and the dynamic environment. It leads to failure of motion planning for intelligent agents like autonomous vehicles. In this paper, we propose a fluid-inspired model to estimate collision risk in road scenes. Multi-object states are detected and tracked, and then a stable fluid model is adopted to construct the risk field. Objects’ state spaces are used as the boundary conditions in the simulation of advection and diffusion processes. We have evaluated our approach on the public KITTI dataset; our model can provide predictions in the cases of misdetection and tracking error caused by occlusion. It proves a promising approach for collision risk assessment in road scenes.
BibTeX
@inproceedings{li2020fluid,
  author = {Xuanpeng Li and Lifeng Zhu and Qifan Xue and Dong Wang and Yongjie Zhang},
  title = {{Fluid-inspired field representation for risk assessment in road scenes}},
  booktitle = {Computational},
  year = {2020},
  doi = {10.1007/s41095-020-0190-8},
}
Figure