Research On Human Action Recognition Based On Joint Model

Human action recognition has an important position in safe operation training, intelligent monitoring and other fields. Dynamic time warping algorithm has been widely used in human action recognition field, which has low computational complexity and strong time robustness. But for the complexity of human action, light and noise, it's difficult to extract features of human action accurately, leading to the low action recognition rate. Therefore, it's necessary to improve traditional action feather extraction method for sequence images, and develop a feather method whit high accuracy, good stability and strong anti-interference.According to the characteristics of human skeleton, slope constraint for burrs cutting has been proposed to solve the wrong judgment caused by the method of length constraint in human skeleton extraction. Trunk endpoints are positioned by chain code traverse method. Search range of traversal structure has been extended to improve the applicability of traditional chain code traversal structure. The human body proportion method for limb joints location was proposed to solve the large deviation caused by the method of maximum triangle area when curvature of limbs was small. For joint coordinates can't accurately express action feather accurately, action feather has been expressed in the weighted normalized vector form.Based on VS2013 and openCV, human skeleton feather extraction and joint positioning in sequence images from Weizmann database was realized. Human action recognition based on DTW algorithm and weighted normalized action feather was realized. Experiment results showed that the methods proposed in the paper could extract action feathers and recognize human action accurately.