<p>LiDAR-based gait recognition is often constrained by the severe trade-off between structural fidelity and inference efficiency. We present MGHP (Multi-Granularity Hierarchical Pyramid), a lightweight framework designed to address this challenge by leveraging depth maps as a compact yet semantically rich representation. Unlike conventional projection-based methods that suffer from semantic fragmentation and the loss of spatial information, MGHP constructs dynamic multi-scale spatial units directly from depth maps to capture hierarchical feature representations. This architecture effectively balances global structural priors with fine-grained micro-motions. To further boost discriminability, we introduce a Metric-Dominated Optimization (MDO) strategy integrated with a convergence-aware scheduling to explicitly reshape the complex feature manifold. Experimental results on the SUSTech1K dataset demonstrate that MGHP attains a Rank-1 accuracy of 87.4% and a Rank-5 accuracy of 96.3%. Relative to state-of-the-art point-based models, MGHP achieves an inference throughput of 878.61 FPS—representing a 17.5-fold speedup. These results indicate that depth-map-based hierarchical modeling is a highly effective solution for high-performance gait recognition on resource-constrained edge devices.</p>