Zero-shot skeleton-based action recognition method based on semantic-to-parameter generation

By using a semantic-to-parameter generation method and leveraging a diffusion model and weight mapping branches, the efficiency and robustness issues of unseen categories in zero-shot skeletal motion recognition are addressed. This enables accurate recognition and adaptive decision-making for new motion categories, improving the flexibility and accuracy of the recognition system.

CN122196687APending Publication Date: 2026-06-12XIDIAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
XIDIAN UNIV
Filing Date
2026-03-20
Publication Date
2026-06-12

AI Technical Summary

⚠Technical Problem

Existing zero-shot skeletal motion recognition methods suffer from limitations in recognition efficiency and robustness when faced with unseen categories. This is mainly due to geometric bias caused by incomplete training data coverage and cross-modal semantic gaps, making it difficult to achieve accurate recognition of unseen categories.

⚗Method used

A semantic-to-parameter generation method is adopted, which utilizes the diffusion model and weight mapping branch in the pre-trained zero-shot skeleton behavior recognition model. Through parameter generation of text adapter and skeleton adapter, dynamic mapping and matching of text semantics to skeleton features are realized, and independent decision boundaries are constructed.

🎯Benefits of technology

It significantly improves the model's generalization ability and robustness to unseen actions, enabling it to adapt to new action categories without retraining, achieve dynamic adaptive decision-making, and expand the application scope of the recognition system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122196687A_ABST
    Figure CN122196687A_ABST
Patent Text Reader

Abstract

The application discloses a zero-shot skeleton behavior recognition method based on semantic-to-parameter generation, converts the decision process from feature comparison to parameter instantiation, explicitly constructs an independent decision boundary for each new category, overcomes the geometric deviation between domains, significantly improves the generalization ability of the model to unseen actions, and greatly improves the recognition accuracy of the model when facing complex actions that have never been seen before; in combination with a text adapter and a skeleton adapter, two modalities, not only effectively alleviate the heterogeneity between modalities, but also ensure that semantic knowledge can be more accurately and stably transferred to skeletal motion representation, significantly improving the robustness and consistency of the model in cross-modal reasoning. The application proposes a semantic-to-parameter generation framework, aiming to shift from feature matching to parameter instantiation and introduce a robust knowledge transfer mechanism. It not only improves the action recognition efficiency in the zero-shot scene, but also provides more objective and scientific support for subsequent behavior analysis.
Need to check novelty before this filing date? Find Prior Art