A little more digging turned up a recent paper that uses Histogram of Oriented Gradients (HOG) along with neural-network training: http://www.jsoftware.us/vol13/355-SE3002.pdf