The invention relates to an action evaluation method based on
human body three-dimensional articulation point detection, which belongs to the field of
computer vision and comprises the following steps: S1, performing
human body three-dimensional articulation point detection on a single-frame picture after video framing; S2, extracting key frames of a specified number of frames of the video; S3, constructing
motion vector features and joint
kinetic energy features, and extracting feature values; and S4, constructing a
key frame action similarity comparison model through multi-
feature fusion: fusing the sub-features in the step S3, and constructing a personalized model for different types of actions; constructing a
motion vector feature similarity function based on
cosine similarity, and constructing a joint
kinetic energy similarity function based on
a weighting function; and obtaining a
key frame action similarity comparison model based on the two similarity functions, comparing the action to be detected with the
key frame set of the standard action, and finally obtaining the action similarity of the motion video. The method is more accurate and scientific, and can be used for
physical fitness action correction and teaching.