GIAT-SymNet: Neuro-symbolic Learning for Explainable and Robust Gait-based Anthropometric Recognition

This paper introduces GIAT-SymNet, a gait-oriented recognition model that blends trainable visual encoders with explicit anthropometric rules so that the system remains both accurate and explainable in non-ideal conditions.First, a convolutional spatio-temporal backbone extracts gait descriptors from silhouette or pose sequences.Next, a symbolic layer evaluates these descriptors through human-readable constraints such as stride-height proportion, left-right swing balance, and torso steadiness, and feeds back rule-consistency penalties to the deep part of the network.This coupling makes the final prediction traceable to the rules that fired.We assessed the model on CASIA-B, OU-ISIR, and a GIAT-enriched set that contains additional body-proportion information.GIAT-SymNet reached 96-97% rank-1 recognition and a macro-F1 above 0.96 under our evaluation protocol, outperforming plain CNN/LSTM baselines, and the performance drop under occlusion and cross-view testing remained below 10% across the reported settings.More than 80% of decisions could be linked to activated rule chains, showing that the added symbolic stage actually improves transparency rather than just reporting saliency.To support reproducibility, the training split protocol, baseline re-implementation settings, and statistical reporting (mean/variance across runs) are provided in the Experimental Setup and Results sections.The results indicate that inserting structured, anthropometry-aware reasoning inside a deep gait pipeline yields a recognizer that is competitive, robust, and auditable, and therefore suitable for publication in vision/biometrics venues that require interpretability.