Human body action recognition method

Abstract

The invention provides a human body action recognition method. The method comprises the following steps of extracting movement of each skeleton joint point of an action sample between adjacent framesto serve as dynamic characteristics of the action sample; performing spatial multi-scale division on the dynamic features to obtain each sub-feature set; for each sub-feature set, enabling motion features of all skeleton joint points in the same frame to form vectors; extracting frame feature vectors of the sub-feature sets of all training samples, and performing clustering to obtain a clusteringcenter; inputting the feature vectors of all frames of the action sample into probability distribution neurons constructed by each sub-feature set, and accumulating all outputs on each probability distribution neuron to obtain histogram expression; performing time multi-scale division on the sub-feature set to obtain a time multi-scale histogram; forming a space-time multi-scale soft quantizationhistogram; forming a space-time multi-scale soft quantization network; and training the space-time multi-scale soft network, and inputting the test sample into the trained network model to realize action recognition.

Description

technical field [0001] The invention relates to a human body action recognition method, which belongs to the technical field of human body action recognition. Background technique [0002] Human action recognition is an important research direction in the field of machine vision, and has a wide range of applications in the fields of human-computer interaction, virtual reality, video retrieval, and security monitoring. With the development of depth cameras and human skeleton extraction algorithms, people can easily obtain information about human skeleton joints. Since the human body can be regarded as a system constructed by the interconnection of rigid skeletal joints, action recognition based on human skeletal joints has significant advantages over image-based action recognition. [0003] In recent years, many methods based on clustering and statistical models have been proposed for action recognition tasks. The codebook is obtained by clustering the features of all sampl...

AI Technical Summary

Problems solved by technology

However, the traditional hard quantization method is difficult to measure and optimize the clustering results, and the features can only belong to a single cluster center, which makes the action description not flexible enough

In addition, the clustering and histogram quantization processes are separated into two stages, making this type of method not end-to-end, and the training process cannot be supervised by sample labels, which also reduces the effectiveness and discrimination of action descriptions.

Images