<p>Mapless autonomous navigation in constrained pedestrian environments requires mobile robots to transform local semantic and geometric perception into physically executable motion commands. This paper presents a semantic-geometric fusion navigation framework for a quadruped robot operating on tactile paving with static and dynamic obstacles. A morphology-based topological reconstruction module converts visual semantic masks into continuous centerline references without relying on pre-built maps, while also improving robustness to local fractures and noise. Visual tactile-paving trajectories and LiDAR obstacle observations are then projected into a unified robot body frame, providing a metric representation for local planning. Based on this representation, a task-specific nonlinear model predictive control (NMPC) scheme integrates semantic path tracking, geometric obstacle avoidance, angular-velocity response approximation, command smoothing, and execution-layer safety filtering. Experiments on a Unitree Go2 quadruped robot equipped with RGB-D and LiDAR sensors demonstrate improved path coverage, gap-repair accuracy, collision-free traversal, and reduced command oscillation compared with conventional path extraction and local planning baselines. The results show that the proposed framework improves the robustness and executability of quadruped navigation in tactile-paving environments.</p>