数据-知识混合驱动的多部位特征共享步态识别方法OA

To address the critical challenge of compromising personnel identity authentication in intelligent factory surveillance,caused by severe occlusion of traditional biometric features(e.g.,face)due to mandated safety protocols(safety helmets and workwear),we propose a Multi-part Feature-sharing Gait recognition model(MFS-GaitPart)driven by a hybrid data-knowledge framework.At the data-driven level,a Frame-level Triple Feature Extractor(FTFE)captures fine-grained spatio-temporal features of the head,torso,and lower limbs,optimizing local motion representation.At the knowledge-driven level,in-dustrial safety protocols and biomechanical priors are integrated to dynamically constrain feature distributions.The data-knowledge hybrid driving mechanism is realized through the Horizontal Pooling Sharing(HP-S)module.The gait feature parti-tion is guided by industrial knowledge,and the shared weight is constructed based on the biomechanical prior to encode the multi-site collaborative relationship,and the adaptive fusion of data and knowledge features is completed through dynamic re-mapping.Extensive experiments demonstrate that in identity authentication scenarios(standard wear/carry distractions/abnormal intrusion),our model achieves an average recognition rate of 97.48%on the ISGD dataset—outperforming state-of-the-art meth-ods by 0.97 percentage points.On the CASIA-B dataset,it attains an average Rank-1 accuracy of 87.44%,surpassing existing approaches by 1.37percentage points~7.21percentage points.This work provides a high-robustness solution for personnel authen-tication and safety management in smart factories.