Real-time body movement recognition method based on human body key point thermodynamic diagram

Abstract

The invention discloses a real-time limb movement recognition method based on a human body key point thermodynamic diagram, and the method comprises the steps: obtaining an original movement video, inputting the original video into an openpose neural network, obtaining a thermodynamic diagram of a human body in a video frame, and obtaining a plurality of human body key points; finding out a predicted maximum index on the thermodynamic diagram of the channel corresponding to each human body key point, and taking the maximum index as a center; a two-dimensional Gaussian distribution function with the standard deviation of 0.6 is adopted, the maximum index in the second step serves as the center, the product of the point and eight points around the point and the two-dimensional Gaussian distribution function is calculated, and a thermodynamic diagram corresponding to a channel is generated; stacking the thermodynamic diagrams obtained in the step 3 along a time dimension, and sending the stacked thermodynamic diagrams into a 3D convolutional neural network for training; and performing action prediction on the trained network to obtain a network output category result, and displaying the network output category result. The method has the characteristics of simple data processing steps and low calculation amount, and can improve the behavior recognition efficiency on the basis of ensuring the precision.

Description

technical field [0001] The invention relates to the technical fields of image and video processing and motion recognition, in particular to a real-time body motion recognition method based on a heat map of key points of the human body. Background technique [0002] With the increase in the deployment of surveillance cameras, action recognition is an emerging application requirement. This method does not need to rely on surveillance personnel to observe surveillance videos for a long time, but directly outputs surveillance cameras or videos in an intelligent recognition method. The target limb action label in . The prevalent methods are RGB image frame-based methods and human keypoint detection-based methods. The problem of the former is that it consumes too much computing resources, and it is easy to mix irrelevant background information, and it is impossible to focus on the target person itself; the problem of the latter is that the efficiency of generating key points of t...

AI Technical Summary

Problems solved by technology

The problem of the former is that it consumes too much computing resources, and it is easy to mix irrelevant background information, so it is impossible to focus on the target person itself; the problem of the latter is that the generation of key points of the human body is inefficient, and the generation of prediction information contains After the heat map, complex post-processing is required to calculate the more accurate position of the key points of the human body

Images