Autonomous driving requires advanced technology to implement completely self-driving vehicles. Perceptual technologies, such as object detection, segmentation, and depth estimation, have been proposed by applying deep neural networks. Segmentation technology plays a crucial role in securing safety by indicating drivable areas. However, it is not sufficient to indicate only the road area to avoid potential danger, as there are various detailed elements that need to be identified even in the road area. Thus, it is necessary to indicate drivable areas that are close to the driver's cognition. Detailed segmentation is important for risk assessment and is a crucial component of our project. Since the expressible range differs according to the depth of a deep neural network, the feature values extracted in the network differ depending on the target object. To address this, we propose to extract and fuse various features and utilize them according to the target objects. We designed the proposed method to generate more detailed segmentation results by adding data labeled with additional classes to the existing large-scale dataset. Our approach is a driver's cognition-based semantic segmentation by multi-level feature extraction, which we call DCSeg-Net for short.