Model gait has the characteristics of stable posture rhythm, limb coordination and trajectory control, which can be expressed by structured posture sequence. This paper proposes a framework of posture feature learning and spatio-temporal relation encoding for modeling gait style of models. The dataset contains 4680 gait sequences of six categories of styles, of which 3276 samples are used for training, 468 samples for validation, and 936 samples for testing. After normalization and segment segmentation, the key point flow is mapped into a time-aligned pose map, and input into a hierarchical network that integrates local joint interaction, spatial cooperative propagation and stage gating mechanism for encoding. Experimental results show that the accuracy of the method reaches 94.6%, Macro-F1 reaches 93.4%, feature clustering purity reaches 91.7%, and the average inference time is 1.21s. The framework can support gait analysis and interpretable style representation of models, and serve fashion action understanding and retrieval in computer vision.
