Research On Human Action Recognition Based On Depth Maps And Linear Representation Model

Because of its application in the future such as intelligent monitoring,intelligent transportation,human action recognition has become an important research direction of computer vision.The early human action recognition research mainly focuses on the traditional RGB data and many classic algorithms have been put forward and applied.However,the traditional RGB data is easily disturbed by the external environment,progress in human action recognition research development is slow.With the development of the technology,Microsoft has released depth image collecting equipment(Kinect),provides new impetus for the development of human action recognition technology.The depth data not only contains the spatial information but also are not easily influenced by illumination change.Thus the research on action recognition based on depth data becomes popular.This paper mainly studies the human action recognition based on depth maps and linear representation model.The concrete contents are as follows:(1)This paper provides a human action recognition method based on discriminative collaborative representation classifier.Firstly,the depth motion maps(DMMs)are extracted from the depth image sequences,which not only are not easily influenced by the environment but also can describe the physical characteristics of the object.Secondly,this paper presents a discriminative collaborative representation classifier(DCRC)based on sparse representation classifier.It can improve the performance and robustness of classifier with a quadratic constraints on coefficients.Finally,DCRC is used to classify DMMs to realize human action recognition.(2)Because DMMs may lose some temporal information and make it difficult to recognition two similar actions with reverse temporal orders.So DMMs pyramid(DMMP)feature descriptor is proposed.This descriptor contains different scales of DMMs,making the feature descriptor more abundant.Then discriminative collaborative representation classifier is applied to classify the extracted features.(3)This paper proposes a method of action recognition based on meta-Action dictionary learning.Since the sparse representation classifier uses training samples as an over-complete dictionary directly and contains too much redundant information,a meta-Action dictionary learning classifier is proposed based on meta-genes and dictionary learning.At the same time,we put forward a weight DMMs pyramid(WDMMP).Compared with DMMP,WDMMP contains more accurate information.Finally,the meta-Action dictionary learning classifier is used to classify the extracted features.