Parameter selection of Gaussian kernel SVM based on local density of training set