直接上代码：

#include <opencv2/opencv.hpp>
using namespace cv;

#include <cmath>
using namespace std;

template<typename T>
int _get_hog(vector<float> &hist, const Mat &img, const int nbins = 8, const bool need_normalize=false){

  int rows_minus_1 = img.rows - 1;
  int cols_minus_1 = img.cols - 1;

  float dx, dy;
  float angle;

  const float angle_base = atan2f(0, -1);
  const float angle_piece = 2.f*angle_base / (float)nbins;

  int bin_id;

  for (int y = 0; y < rows_minus_1; y++){
    for (int x = 0; x < cols_minus_1; x++){
      dx = (float)(img.at<T>(y, x) - img.at<T>(y, x + 1));
      dy = (float)(img.at<T>(y, x) - img.at<T>(y + 1, x));

      angle = atan2f(dy, dx) + angle_base;

      bin_id = (int)floorf(angle / angle_piece);

      hist[bin_id] += 1.f;
    }
  }

  hist[nbins - 1] += hist[nbins];
  hist.resize(nbins);

  if (!need_normalize){
    return 0;
  }

  float num_pixels = (float)(rows_minus_1*cols_minus_1);

  for (int i = 0; i < nbins; i++){
    hist[i] = hist[i] / num_pixels;
  }

  return 0;
}

// support
// CV_8UC1
// CV_32FC1
// CV_64FC1
int get_hog(vector<float> &hist, const Mat &img, const int nbins = 8, const bool need_normalize = false){
  if (img.type() != CV_8UC1 && img.type() != CV_32FC1 && img.type() != CV_64FC1){
    cerr << __FUNCDNAME__ << " invalid image type!" << endl;
    return 1;
  }

  hist.resize(nbins+1);
  hist.assign(nbins+1, 0);

  if (img.type()      == CV_8UC1){
    return _get_hog<uchar> (hist, img, nbins, need_normalize);
  }
  else if (img.type() == CV_32FC1) {
    return _get_hog<float> (hist, img, nbins, need_normalize);
  }
  else if (img.type() == CV_64FC1) {
    return _get_hog<double>(hist, img, nbins, need_normalize);
  }

  return 1;
}