OpenCV cascades with HOG are only available in 2.4.x. That means you are basically using an outdated and old version of the functionality. I advise against that.
OpenCV does not support adding multiple features into a single cascade. For that we would need to change the backend and allow any arbitrary number of features for a model to be trained. I still think this is a valid idea but we need someone with time to implement it. If you would be up for it, you could suggest it as a GSoC2018 project proposal!
