A novel PCA (principal component analysis) classifier is presented, which calculates the principal components of each class and designs the classifier according to the projection of the data on the subspaces spanned by these principal components corresponding to different classes. It is suited especially for cases where the data of different classes are distributed in different styles and different directions. It can also be easily applied to multi-class problems. Experiments on artificial and real data sets showed its advantage over some other classifiers, such as Fisher discriminant and linear support vector machine. The PCA classifier is applied in the practical problem of vehicle recognition and detection from static images.