<p> Images captured in road scenes are often degraded by atmospheric particles such as fog or haze. This degradation adversely affects the performance of vision-based perception systems, including object detection and stereo matching. Many conventional dehazing methods are based on the atmospheric scattering model, in which the global atmospheric light and the transmission map are estimated. However, estimating these parameters from a single image is inherently ill-posed and remains challenging. To address this, we propose DANet, a depth-aware convolutional neural network that takes an RGB image along with its corresponding depth map as input. In practice, depth maps obtained from hazy images often contain errors, and thus relying solely on them for dehazing may not be reliable. Therefore, DANet learns jointly from RGB and depth information to mitigate the impact of such errors, constrain the solution space of the transmission map, and enhance restoration quality. Experimental results demonstrate that DANet not only performs well on synthetic hazy images used for training but also generalizes effectively to real-world hazy road images beyond the training data. We evaluated DANet on images captured in real driving scenarios and the publicly available DrivingStereo dataset, showing superior dehazing performance compared to existing methods.</p>