The depth estimation of foggy images has always been a major challenge in the research field. The common depth estimation methods use supervised training to estimate the depth of foggy images, but their effectiveness is often limited by the domain adaptation characteristics of supervised training. Here, we propose an unsupervised domain separation depth estimation algorithm for foggy images. This algorithm adopts an unsupervised approach and designs a domain separation framework for foggy and clear images to perform depth estimation on foggy images. It utilizes the characteristic that depth information can be used for both dehazing and hazing, incorporating a self-depth domain conversion module that constructs a symmetric training framework. Domain separation separates the information of the image itself in the feature space dimension, breaking it down into two parts: exclusive domain information (color, lighting, fog degree, etc.) and common domain information (depth information). The experimental results show that our designed network can achieve state-of-the-art results on the NYUv2 dataset and SUN RGB-D dataset, which is superior to existing advanced depth estimation algorithms. Furthermore, the algorithm has strong robustness and can accurately estimate the corresponding depth maps for both non-foggy and foggy images.