Action recognition method based on human body joint points

Abstract

The invention relates to action recognition, in particular to an action recognition method based on human body joint points, which comprises the following steps: acquiring an action behavior video, extracting image features and image convolution kernels from a frame picture, and acquiring a human body joint point sequence corresponding to the frame picture based on human body skeleton information; inputting the image features and the human body joint point sequence into a graph convolutional neural network to obtain a first action recognition result; obtaining a spatial-temporal feature vector based on the graph convolution kernel and the human body joint point sequence, and obtaining a second action recognition result according to the spatial-temporal feature vector; analyzing and judging the logic matching degree between the first action recognition result and the human body posture in the corresponding frame picture as well as between the second action recognition result and the human body posture in the corresponding frame picture, and determining a final action recognition result; according to the technical scheme provided by the invention, the defect of poor accuracy of human body action recognition in the prior art can be effectively overcome.

Description

technical field [0001] The invention relates to action recognition, in particular to an action recognition method based on human joint points. Background technique [0002] The research on human action recognition has important scientific significance in many fields such as computer video surveillance and human-computer interaction. One of the purposes of action recognition is to provide user-related behavior information and use computer systems to actively assist users in completing tasks. Human action recognition can be used for natural human-computer interaction, and is widely used in many fields such as motion analysis, virtual reality, augmented reality, education, film and television animation production, etc. [0003] Human action recognition began in the late 1990s, and the research methods can be roughly divided into two categories: vision-based methods and sensor-based methods. With the emergence of RGB-D sensor devices such as Kinect, the research on human actio...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the above-mentioned shortcomings existing in the prior art, the present invention provides an action recognition method based on human joint points, which can effectively overcome the defect of poor accuracy of human action identification existing in the prior art

Images