Estimating Fog Parameters From a Sequence of Stereo Images

Area of Science:

• Computer Vision
• Robotics
• Environmental Sensing

Background:

• Existing methods for estimating fog parameters from stereo images often suffer from error propagation due to sequential parameter estimation.
• Real-world fog is typically globally inhomogeneous, posing challenges for algorithms that assume local homogeneity.

Purpose of the Study:

• To develop a novel method for simultaneously estimating fog model parameters from stereo foggy images.
• To create a comprehensive dataset for evaluating visual perception algorithms in foggy conditions.
• To improve the robustness and accuracy of visual Simultaneous Localisation and Mapping (SLAM) and odometry systems in fog.

Main Methods:

• A novel optimization problem is formulated to estimate all fog parameters simultaneously, overcoming limitations of sequential approaches.
• The method assumes local homogeneity of fog, enabling effective handling of globally inhomogeneous real-world fog.
• Development of the Stereo Driving In Real Fog (SDIRF) dataset, featuring calibrated stereo frames of real foggy road scenes and corresponding clear-weather data.

Main Results:

• The proposed algorithm demonstrates superior performance compared to prior methods on both synthetic and real foggy data.
• Simultaneous estimation of fog parameters leads to more accurate results and better adaptation to real-world fog conditions.
• The SDIRF dataset provides essential calibrated photometric parameters for accurate atmospheric scattering model application.

Conclusions:

• The developed method offers a significant advancement for visual perception in foggy environments.
• The algorithm can be seamlessly integrated as an add-on module into existing visual SLAM and odometry systems.
• The public release of the code and SDIRF dataset aims to foster further research in fog-affected visual perception.