One of the most popular tasks in the computer vision discipline utilizing deep learning techniques is semantic segmentation. However, visual understanding of complex urban street scenes may be challenging regarding both correctly semantically classifying innate imbalanced datasets and supporting mechanisms for making robust neural network architectures against image quality degradation and generalization on inference time. In this paper, both of these challenges are addressed on the Cityscapes dataset by providing solutions in terms of data augmentation, external datasets, transfer learning, and penalization class weights techniques on a U-Net-based architecture. We evaluate these mechanisms, firstly, by implementing an off-the-shelf U-Net architecture serving as a baseline model, and secondly, by making slight adjustments to that architecture to create an ensemble model comprising various decoders where our proposed techniques are applied during the training phase. The results show that our solution achieves satisfactory performance when being evaluated on the Cityscapes dataset as well as on the internal challenge of the CodaLab competition server.