Our investigation aims at constructing random oblique decision trees to recognize handwritten digits. At the pre-processing step, we propose to use the GIST descriptors to represent digit images in large number of dimensions. And then we propose a multi-class version of random oblique decision trees based on the linear discriminant analysis that is suited for classifying high dimensional datasets. The experimental results on USPS, MNIST datasets show that our proposal has very high accuracy compared to state-of-the-art algorithms.