Robust Dense Depth Estimation in Foggy Weather Conditions

Abstract

This article addresses the challenge of obtaining reliable dense depth maps from outdoor stereo images captured in foggy weather conditions, which can help vehicles to achieve auto-driving in foggy weather. Traditional methods often fail to account for the impact of fog. Deep learning approaches face difficulties due to the randomness in the type and density of fog, making each fog event unique and training models effectively challenging. To better solve the stereo-matching problem in dense fog conditions, we propose a novel method that leverages preserved information in the matching cost function and neighboring disparity values. By utilizing functional analysis on the matching cost function, we generate candidate disparity results, which are filtered based on neighborhood information. Finally, we decouple the overall energy function to construct univariate functions for each pixel, obtaining the final disparity results through minimization. Experimental evaluations demonstrate the effectiveness of our method in achieving more accurate disparity results in foggy weather.