I would suggest that high dimension features are often hard to train against, and hard to validate. I would suggest that whichever your discriminant and training approach, you may want to consider reducing the dimentionality of training either by K-L or principal components transforms, or by a directed layer architecture (identifying subclasses, which may be combined for your final classification). In both instances the intent would be to reduce the variation each element of the system need be tested against.

Think in terms of a training set of size proportional to #classes^#features. If you can break the problem down, by either reduction of features or by logical reduction of classes, training and validation become much more tractable.

First I suggest making a selection of variables or the same to reduce the dimensionality of the features (attributes) of classes. this can be done with principal component analysis, independent component analysis, testors (BT algorithm) typical, etc.. there are tools like weka, R, Matlab, etc.. to see the behavior of your data, there can understand more the behavior of the data you want to analyze.

And then do the clasisficación with k-nn, fisher linear discriminate, multiclass SVM, etc.. You can use the opencv library with java (Eclipse, Netbeans, etc.) Or C (Mono, Codeblocks, etc..), Android with opencv (Adroid studio, etc.), etc. ..