Research On Emotion Recognition Of Body Motions Based On The Shape Of Gaussian

Body motion includes a lot of a ect information, is becoming a new area of a ectrecognition. At present, existing emotion recognition based on the body motions worksare too weak, mainly realized by human subjects judging or taking basic physicalfeatures with various machine learning techniques, have low recognition rates. To solvethis problem, it's necessary to utilize strong features which could quantify motions ofdifferent emotion classes accurately, and combine machine learning to achieve higherefficiency.The existing features are mean of common motion information (speed, acceleration,jerk etc.), ignoring the space characteristic of body movement. How to extract features topicture motions completely? We take consideration of shape of Gaussian (SOG) inmathematics, which is proposed for that covariance cannot form a completeparameterized multivariate Gaussian without the mean parameter. Besides inheriting theadvantages of mean, it has the ability of depicting space information as a probabilitydensity function. So we describe human body motion by SOG to improve the wholerecognition rates. Though a SOG represents a corresponding single Gaussian model,which has less flexibility than Gaussian mixture model clearly in application, the latter isextension and advancement on the single Gaussian model. Thus based on the SOG, weintroduce mixture Gaussian feature to describe different emotions more completely.Through observing the feature vector, the features extracted have nonlinear relation withemotion classes, so we take machine learning technology to recognize the emotioneffectively. Based on different datasets, different kernel functions, different featurecombinations, we conduct N fold cross validation tests. Dividing the samples into N partswith an equal proportion, we take one part as unknown samples to run N cycle tests. Foreach iteration, we use support vector machines to train the known samples marked withemotions and establish a model, then utilize the model to classify the unknown samples. Finally we get the all tests'recognition results.Analyzing the results from the performances of kernel functions, the description capability and the popularity capability of features etc., Experiments show better recognizing rates with SOG or mixture Gaussian feature compared with common feature. Furthermore, mixture Gaussian feature performs better than SOG in description capability, but the latter has more strong popularity capability than the former.