Sensor-Based Gait Recognition Using a Regularized Attention Model to Reduce Covariates

Gait recognition using sensor signals has gained prominence as a non-intrusive biometric technique, offering robust applicability in diverse real-world settings. However, existing models face significant challenges: (i) they inadequately model temporal variations in gait cycles due to covariates such as clothing, surface, and speed, hindering generalization across subjects and conditions, and (ii) they struggle to jointly capture short- and long-term temporal dependencies in sequential gait signals. To overcome these limitations, a Regularized Channel Attention (ReChAtt) framework is introduced, which adaptively enhances channel selection by integrating temporal awareness and regularization. The framework employs a Maximum Entropy Discriminator to promote broader utilization of informative channels, addressing covariate-induced variations, while a Kullback–Leibler divergence regularizer aligns attention weights with class-discriminative distributions, mitigating overfitting and enhancing generalization. ReChAtt is seamlessly integrated into a Multi-kernel Temporal Convolutional Network backbone, augmented with positional encoding to capture multi-scale temporal dependencies and cycle-phase awareness effectively. Extensive evaluations of benchmark gait datasets (OU-ISIR, whuGAIT, and IDNet) demonstrate competitive state-of-the-art performance, achieving accuracies ranging from 89.62 % to 99.72 % . The proposed approach yields consistent improvements over conventional attention mechanisms across multiple backbone architectures and remains competitive with existing methods, particularly on long-sequence and covariate-rich settings.