I highly recommend you to read "Pedestrian Detection: An Evaluation of the State of the Art" by Dollar et al. They've done an elaborate study on human detection back in 2012 on a larger and better dataset (publicly available).
HOG is not state of the art in human detection anymore and the algorithms evaluated in this paper are probably not state of the art too.
But this paper will lead you to state of the art methods. I recommend you to look at papers (from good conferences like cvpr, iccv, bmvc, ... and good journals like IEEE transactions, PAMI, ...) that have cited this evaluation paper.
I have re-implemented it in C++, Matlab, Python and CUDA. However it is not a public implementation. You can find a public implementation in C++ and another in CUDA in the OpenCV library. However, it only has the detection part and not the training part that can be very tricky. However, it is already trained for many object classes: