Research Of Human Action Recognition Based On Global Information

Human action analysis has been a widely studied field,which is a comprehensive application of computer vision machine learning and other disciplines.Previous studies are based on RGB images,but methods' performance is sharply affected by illumination,occlusion,and perspective changes.Meanwhile,methods based on depth camera can't only provide 3D skeleton joint position,but also they are slightly affected by illumination viewpoint change and other external environment.With the popularity of depth camera and technology assessing skeleton joint position from depth image matured,human action analysis of 3D skeleton data has gradually become a new hotspot in the field.Human action analysis contains two important parts: action segmentation and action recognition.The meaning of action segmentation is dividing the motion sequence into several segments according to the class attribute.The meaning of action recognition is recognizing action category,such as running,walking,jumping et al.Because of the influence by the external environment,individual difference,the complexity of action and the equipment precision degree,human action analysis is still a challenge work.For human action segmentation,based on the existing segmentation and thorough analysis,adaptive segmentation Approach is proposed,which is an extension of change point detection algorithm and clustering algorithm.In fact,the species of action is unknown in given action sequence.However,it is the premise for clustering algorithm.On the other hand,change point detection algorithm can only detect transition point of adjacent and different action clips.The same specie can't be distinguished when they are not adjacent.So the change point detection belongs to rough segmentation for action sequence.For the above problem,the proposed algorithm tries to give a feasible solution.Firstly,we use change point detection algorithm to divide action sequence into several clips.Secondly,maximum mean difference is utilized to measure the similarity between clips and corresponding similarity matrix is established,and then the species of action are computed by principle component analysis.Finally,clustering algorithm is used to optimal segmentation results.For human action recognition,Global Gaussian weighted feature representationis proposed,which is another work of this paper.The skeletal data captured by depth camera contains a lot of noise.Previous feature representation which focuses on local feature,cannot achieve satisfactory recognition accuracy.Due to the continuity and smoothness of human action,Global Gaussian weighted feature representation can get a better result on the inhibition of noise.The proposed algorithm establishes independent Gaussian models based on every joint's Euler angle along the same direction of the global action sequence.Then the paper uses Gaussian weighted of the Euler angle as feature representation.Furthermore,the paper uses Hidden Markov model to train and recognize human action data.Based on two improvements,human action analysis has achieved good performance on two different benchmark datasets and our dataset.