Posture Recognition In Static Background

The research of posture recognition of the human body in the video is combined with the research of computer vision and pattern recognition.At present,there are many researches about human posture recognition.While most of them are based on the video data captured by technical equipment,such as Kinnect,human motion sensor and depth image devices.Locating the joints and building the human skeleton directly by the technical equipment to analyze and recognize human posture.In applications,it is not possible to complete the recognition in video captured by technical device.It is more common to obtain and analyze human postures in video captured by universal equipment.For example,security monitoring,the identification of the human's movement in the surveillance video is useful for that timely security alarms and predictive protection and other related measures can be taken.While it is relatively difficult to extract human skeletons from ordinary videos.Generally,the target's regional characteristics are used to distinguish human postures,and human body characteristics are not fully utilized,and the recognition rate is not high enough.Based on the video captured by universal equipment,this thesis researched the posture recognition and behavior detection of single-target humans in static scenes.The static scene means that the shooting equipment is static and there is no moving object in the background,while there may be moving things such as leaf shaking,light shadow,etc.The single target human postures include stand and bent,and the single target behavior actions include walking,running,bending and getting up.In the related researches,there are many problems that targets are inaccurate,skeleton extraction is difficult,and recognition rate is low.The innovations in this thesis include:(1)For the target's shadow removal,a novel method based on the center point of the target area,Shadow Removal Based on Central-point,is proposed.With the shadow area is far from the center point of the target and the brightness of shadow is lower than target,the shadow interference area is removed.Compared with other algorithms,the new method has removed shadow area more accurately.(2)It improved the skeleton extraction method,base on two used methods of skeleton extraction,morphological skeleton extraction and the skeleton extraction by pixel layer's domain computing.A continuous and complete skeleton,instead of full of redundant parts by the former and incomplete by the latter,is obtained.(3)A method that make skeleton to be single pixel based on direction trend,Dtrends(Direction trends)method,is proposed.single pixelized human skeleton is obtained by judging deal the trends of the skeleton points,and it avoid the serrated shape of the edge of the eight-neighbor single pixel method.With related researches and above innovative works in this thesis,a single pixelized human skeleton is obtained.Analyzed the standing and bending posture features,a set of two-dimensional data that used to classify the human posture by SVM is extracted from the human skeleton model.Combined with the analysis of the human posture and the movement of the body centroid,the behavioral movement detection is completed.After compared the results of related researches,this method has achieved a higher posture recognition rate and behavior detection correct rate.