Research Of Human Behavior Recognition Method

Human action recognition is one of important hot topics in the field of computervision. The purpose of the study human behavior recognition is to let the machineunderstand the people, and help people do some tasks. It can also help people betterunderstand their kinetic characteristic. At present, human behavior recognition havebroad application prospects in intelligent monitoring and human-computer interactioninterface, motion analysis and other fields.Human behavior recognition is generally divided into motion detection, featureextraction and behavior recognition three processes, the motion foreground in imagesequence is extracted by motion detection, and then features is extracted fromforeground, which used for behavior classification and recognition. Therefore, thisdissertation mainly research work focuses on these three processes and methods used.Firstly, this dissertation introduces several kinds of frame difference andbackground subtraction methods. Among these methods, weighted cumulativedifference method could extract a more complete human motion foreground than twoframes difference method and three frames difference method, compared with theaverage background model, Gaussian mixture model and Codebook model, weightedcumulative difference method does not need to background modeling and update,therefore it has advantages of simple and fast. But weighted cumulative differencemethod needs to select empirical formula of weights artificially, which will bringinconvenience algorithm design and application. Therefore this dissertation putforward an improved weighted cumulative difference method, through the calculationof the correlation between the frames to determine the corresponding weights, whichwill make it obtain a more complete human motion foreground, and reducing theaccumulation of reducing noise, the effectiveness of the proposed method is provedthrough the behavior recognition experimental results.Secondly, the key frame template selection methods and image blocks methodsare used for behavior representation based on the image area, the proportion of foreground pixels in each block is calculated, then use the proportions to constitute thefeature vectors of the templates. For behavior representation and feature extractionmethod based on the skeleton joints feature, using Kinect to extract the skeleton jointcoordinates of body movement, and the coordinates of the body joints is clustering intoa sequence of symbols through K-means, which is used for human behaviorrecognition based on state space.Finally, this dissertation introduces Bayesian classifier and support vectormachine (SVM), which used for behavior recognition based on template matchingachieved80%and84.44%average recognition rate respectively, which proved thatsupport vector machine (SVM) can obtain good effect when deal with the behaviorrecognition task of small sample, nonlinear and high dimension. For state spacemethod, the dissertation introduces the hidden Markov model, using Kinect obtain theskeleton of joint point coordinates as features for behavior recognition, experimentalresults achieved the average recognition rate of94%.