Action recognition system and method based on human skeleton point motion features

Abstract

This invention discloses a motion recognition system and method based on the motion features of human skeletal points. The method includes: a detection stage: acquiring video stream data, including the number of human figures and feature information of a two-dimensional human skeletal sequence, and constructing a two-dimensional human skeletal motion sequence; a correction stage: correcting the distortion of the two-dimensional human skeleton in the two-dimensional human skeletal sequence to conform to the standard shape under a level view; and a recognition stage: using the corrected two-dimensional human skeletal sequence, extracting the motion feature information of the two-dimensional skeleton, classifying it using a neural network, and completing the motion recognition of the two-dimensional human skeleton. This invention uses a two-dimensional human skeletal sequence to represent the motion of human figures in surveillance videos, which is unaffected by environmental changes and corrects the distortion caused by perspective phenomena. It can adapt to different camera installation angles and also extracts the temporal and spatial dependencies of key skeletal points, effectively representing the motion features of different actions.

Images