Skeleton action recognition method based on prototype strategy in combination with multi-channel space-time diagram convolutional network

By adopting prototype strategies and multi-channel spatiotemporal graph convolution networks in skeleton action recognition, the problems of sample scarcity and multi-scale information fusion are solved, and high-accurate action recognition is achieved under a small amount of data.

CN119694001APending Publication Date: 2025-03-25YUNNAN UNIV
5 Cites 0 Cited by

Patent Information

Application Number
CN202411779266.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-12-05
Publication Date
2025-03-25

AI Technical Summary

Technical Problem

The existing skeleton action recognition methods are difficult to train in the absence of samples, and it is difficult to adapt to multi-scale spatio-temporal information fusion, and it is impossible to accurately identify action sequences with inconsistent lengths.

Method used

The prototype strategy is used to combine the multi-channel spatiotemporal graph convolution network, and the space-time graph of support sets and query sets are constructed, and the feature extraction is used to use the multi-channel spatiotemporal graph convolution network with weight sharing, and the shared space is divided through dynamic time series matching and prototype strategy to identify action categories.

Benefits of technology

The action categories can still be predicted more accurately with a small number of data samples, which improves the accuracy of action recognition and the robustness of the model.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a skeleton action recognition method based on a prototype strategy in combination with a multi-channel space-time diagram convolutional network, solves the problem that a training model recognizes human body actions under the condition of few samples, fuses multi-scale information of a time dimension and a space dimension, and improves the accuracy of skeleton action recognition. The method comprises the following steps: S1, acquiring a to-be-identified skeleton sequence to construct a data set, and dividing the skeleton sequence into a query set and a support set; s2, respectively constructing a support set space-time diagram and a query set space-time diagram according to the skeleton sequences of the support set and the query set; s3, inputting the support set space-time diagram and the query set space-time diagram into the constructed weight-shared multi-channel space-time diagram convolutional network for feature extraction, and generating corresponding support set features and query set features; s4, performing dynamic time sequence matching on the support set features and the query set features to obtain length-aligned sequence features; and S5, according to a prototype strategy, sharing space division and action category identification are carried out on the query set features and the support set features which are aligned in length.
Need to check novelty before this filing date? Find Prior Art