Two-person interaction behavior recognition method based on MSA-CLSTM body local activity

By using the MSA-CLSTM model and employing frame optimization and activity evaluation modules, effective features of two-person interaction behavior are extracted, solving the problems of feature redundancy and balancing action details, and achieving efficient two-person interaction behavior recognition.

CN119832635BActive Publication Date: 2026-06-26ZHENGZHOU UNIVERSITY OF LIGHT INDUSTRY
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
ZHENGZHOU UNIVERSITY OF LIGHT INDUSTRY
Filing Date
2024-12-23
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

Existing methods for recognizing two-person interaction behaviors have failed to effectively address the issue of feature redundancy, leading to increased model burden and difficulty in balancing individual action details with overall interaction attributes, thus affecting recognition performance.

Method used

The MSA-CLSTM model is adopted, and the interactive characters are tracked through frame optimization method. An MSA activity assessment module is established, and individual action details are extracted by two-dimensional discrete wavelet transform. Local skeletal angles and mutual feature flows are calculated, and multiple feature flows are fused to realize the recognition of two-person interactive behavior.

Benefits of technology

It effectively reduces the impact of invalid features on the model, improves the accuracy and efficiency of two-person interaction behavior recognition, and solves the problem of the model focusing too much on individuals or the whole while ignoring details.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119832635B_ABST
    Figure CN119832635B_ABST
Patent Text Reader

Abstract

The application provides a two-person interaction behavior recognition method based on MSA-CLSTM body local activity, first, using pose estimation technology to extract human skeleton key point sequence of interactive characters from an RGB video; through a human body local activity evaluator, key points of the human skeleton sequence are evaluated for activity, a human body local weight is generated, and a key point coordinate feature flow is obtained by weighting the original skeleton sequence; relative distances between human skeleton key points are calculated, a key point relative position feature flow and a mutual feature flow are generated by two-dimensional discrete wavelet transform and a spatial attention mechanism; local skeleton direction feature flows are generated by calculating local skeleton directions from original human skeleton data and calculating angles between skeletons according to a natural connection mode of human skeletons; finally, four feature flows are fused through a multi-feature flow fusion model, two-person interaction action categories are calculated according to the fused features, and two-person interaction behavior recognition is completed.
Need to check novelty before this filing date? Find Prior Art