Urban driving scenes naturally exhibit hierarchical structure (e.g., vehicle \(\rightarrow \) car, truck, bus), yet most perception systems treat categories as flat and rely on costly manual annotations. We present Hier-SimCLR-Drive, a self-supervised framework that simultaneously learns transferable visual representations and a two-level hierarchy of traffic-scene concepts directly from raw, unlabeled video frames. Our method combines contrastive learning—bringing multiple augmented views of the same image closer in feature space—with dual clustering heads that organize encoder outputs into coarse categories (e.g., vehicles, pedestrians, infrastructure) and finer sub-categories (e.g., sedan, SUV, bus, bicycle). This hierarchical organization emerges without human labels, scaling naturally across objects, cities, and environmental conditions. On the Cityscapes dataset, the discovered categories align well with ground-truth labels, and the learned features surpass a flat SimCLR baseline on a frozen linear probe (79.6% vs. 72.2%, +7.4 pp). Beyond improved accuracy, Hier-SimCLR-Drive produces an interpretable, machine-generated taxonomy of road-scene objects, paving the way for more adaptable and transparent perception in autonomous driving.