A Behavior Recognition Method Based on Transformation Module

Abstract

The invention discloses a behavior recognition method based on a transformation module, and relates to the field of human body motion recognition. The behavior recognition method based on the conversion module first reads continuous frame images and constructs a mask; then constructs the input data of the conversion module, including obtaining the input data of the conversion module and the position mask matrix mask operation; and then recognizes the action of the conversion module, including data preprocessing operations After the data processing after the continuous encoding module, the action detection result is obtained; finally, the category detection result is calculated with the category label target to calculate the cross-entropy loss and optimize the network parameters. This method uses the conversion module used in natural language understanding to extract the spatio-temporal features of continuous frame images. At the same time, only the conversion module is used in the entire recognition process, thereby reducing the parameter amount of the method, reducing the overall calculation amount, and improving the performance of the action. Identify the frequency.

Description

technical field [0001] The invention relates to the field of human action recognition, in particular to an action recognition method based on a conversion module. Background technique [0002] Action recognition extracts the action features of continuous video frames, completes the analysis task of video action content and obtains a classification task of action category, which can help improve the monitoring ability of dangerous behaviors in key areas and avoid possible dangerous behaviors. . [0003] The Chinese patent with the patent number CN202010708119.X proposes an efficient unsupervised cross-domain action recognition method (CAFCCN) based on channel fusion and classifier confrontation, which is used to solve the problem that the target data set training data set has no labels. Using the information of the source domain data set and the information of the target domain unlabeled training set, the accurate identification of the target domain test set is realized. The...

AI Technical Summary

Problems solved by technology

The disadvantage of this method is that four deep residual network models are required to extract the characteristics of the optical flow graph in the source optical domain and the optical flow graph in the target domain, and multiple fusion modules are required to complete the characteristics of the optical flow graph in the source optical domain and the target domain. The fusion of optical flow graph features leads to more model parameters of the entire algorithm and a larger overall calculation load

The disadvantage of this method is that in the method, the interval frame images are directly input into the action recognition network for action recognition. When the device is in a complex environment and has multiple targets, the different actions of different targets will affect the action detection results of the entire image, and it is impossible to perform action detection for each target. A target for action recognition

[0008] Compared with 2D convolution, 3D convolution needs to extract continuous frame features in three dimensions, which leads to an increase in the number of parameters of the 3D convolution model, and increases the amount of model calculation, resulting in a longer training period for the model.

At the same time, as a new computing method, 3D has poor support for 3D convolution under different deep learning frameworks, which affects the practical applicability of action recognition algorithms based on 3D convolution.

[0009] However, the optical flow method requires multiple 2D convolution models to cooperate with each other to extract temporal and spatial features, resulting in excessive model parameters and a large amount of calculation. practical applicability

Images