Contrast restoration of road images taken in foggy weather

Houssam Halmaoui, Aurélien Cord, Nicolas HautièreWorkshops @ ICCV 2011 · DOI

Abstract

This paper introduces a hybrid approach for restoring the contrast of road images captured in foggy conditions to improve Advanced Driver Assistance Systems (ADAS). By combining the physical Koschmieder model of light scattering with local histogram equalization and optimizing parameters via simulated annealing, the authors achieved a 12% improvement over existing state-of-the-art dehazing methods.

Introduction

Fog is a major hurdle for camera-based driver assistance systems. It reduces visibility, leading to potential accidents and causing ADAS algorithms—such as lane detection or sign recognition—to fail or trigger false alarms. The core challenge in "dehazing" is that the problem is "ill-posed." According to the standard physical model, a single pixel's intensity depends on four unknowns: the object's true brightness, the sky's brightness, the fog density, and the distance to the object.

The physical Koschmieder model illustrating the components of light scattering in foggy conditions.

Most existing solutions are either too computationally expensive for real-time automotive use or fail to account for the specific geometry of a road, such as the distinction between the flat road surface and vertical obstacles like signs and cars.

Proposed Methodology

The authors propose a two-stage procedure: visibility condition detection followed by contrast restoration. Their restoration method is a unique blend of physical modeling and signal processing, as detailed in the flowchart below:

Flowchart of the proposed contrast restoration procedure, from image input to final restoration.

1. Local Histogram Equalization (LHE)

Instead of solving for depth first, the system uses LHE in the Lab color space to create an initial "rough draft" of the restored image. This enhances local contrast but often introduces "blocky" artifacts.

2. Physical Model Integration

The LHE result is used as an estimate for the intrinsic luminance ($I_0$) in the Koschmieder equation. By rearranging the formula, the system can derive an approximate depth map of the scene.

3. Refinement and Smoothing

To fix the blocky artifacts from the LHE step, Gaussian filters and Sobel edge detection are applied. A "depth map mixing" strategy is employed: the road is assumed to be a flat plane (the "planar assumption") while cars and signs are treated as vertical objects.

4. Optimization

To find the best settings for window sizes and intensity adjustments, the researchers used Simulated Annealing—an iterative optimization algorithm—to minimize the error between restored synthetic images and their "ground truth" fog-free counterparts.

Experimental Setup and Datasets

The study utilized a synthetic database of 66 color images representing various environments (city, highway, rural). Synthetic data is essential for quantitative evaluation because it is nearly impossible to capture the exact same road scene with and without fog in the real world. The researchers tested four variations: uniform fog, variable sky intensity, variable fog density, and a combination of both. Qualitative results were also demonstrated on real grayscale and color images of foggy roads.

Key Results and Insights

  • The Hybrid Advantage: Purely physical models often struggle with unknown depth, while purely signal-based models ignore the fact that fog gets thicker with distance. Combining them allows the signal-based approach to provide a "hint" that helps solve the physical equations.
  • The Planar Assumption: Treating the road surface differently than vertical objects prevents "over-restoration." For example, white lane markings can be mistaken for distant objects; by assuming the road is a flat plane, the algorithm restores these markings more accurately.
  • Quantitative Performance: The proposed method outperformed all compared algorithms, including Multiscale Retinex (MSR) and Dark Channel Prior (DCP). It reduced the mean absolute error by at least 12% compared to the previous best-performing method (NBPC).
  • Qualitative Performance: On real-world images, the method significantly increased the visibility of details such as distant vehicles and road signs without creating unrealistic color shifts.

Conclusion and Future Work

While effective, the current physical model is designed for daytime fog and does not account for artificial light sources or "halo" effects at night. Future work will focus on improving color accuracy by implementing a pre-processing white balance step and integrating this restoration as a pre-processing layer for actual ADAS tasks to measure real-world safety improvements.

Summary updated

BibTeX
@inproceedings{halmaoui2011contrast,
  author = {Houssam Halmaoui and Aurélien Cord and Nicolas Hautière},
  title = {{Contrast restoration of road images taken in foggy weather}},
  booktitle = {Workshops @ ICCV},
  year = {2011},
  doi = {10.1109/iccvw.2011.6130501},
}
Figure