深度图转点云的原理

深度图转点云的计算过程很简洁，而里面的原理实际是从内外参矩阵变换得到．下面来介绍其推导的过程．

1. 原理

首先，要了解下世界坐标到图像的映射过程，可以参考下教程＂相机标定(2)---摄像机标定原理＂，这里不做赘述．用公式具体的表达如下：

\(z_{c}\begin{bmatrix}u\\ v\\ 1\end{bmatrix}=\begin{bmatrix}f/dx & 0 & u_{0}\\ 0 & f/dy & v_{0}\\ 0 & 0 & 1 \end{bmatrix}\begin{bmatrix}R & T\\ 0 & 1\\ \end{bmatrix}\begin{bmatrix}x_{w}\\ y_{w}\\ z_{w}\\ 1\end{bmatrix} \)

2. 代码

#ifndef DEPTH_IMAGE_PROC_DEPTH_CONVERSIONS
#define DEPTH_IMAGE_PROC_DEPTH_CONVERSIONS

#include <sensor_msgs/Image.h>
#include <sensor_msgs/point_cloud2_iterator.h>
#include <image_geometry/pinhole_camera_model.h>
#include "depth_traits.h"

#include <limits>

namespace depth_proc {

typedef sensor_msgs::PointCloud2 PointCloud;

// Handles float or uint16 depths
template<typename T>
void convert(
    const sensor_msgs::ImageConstPtr& depth_msg,
    PointCloud::Ptr& cloud_msg,
    const image_geometry::PinholeCameraModel& model,
    double range_max = 0.0)
{
  // Use correct principal point from calibration
  float center_x = model.cx();//内参矩阵中的图像中心的横坐标u0
  float center_y = model.cy();//内参矩阵中的图像中心的纵坐标v0

  // Combine unit conversion (if necessary) with scaling by focal length for computing (X,Y)
  double unit_scaling = DepthTraits<T>::toMeters( T(1) );//如果深度数据是毫米单位的，结果将会为0.001;如果深度数据是米单位的，结果将会为1;
  float constant_x = unit_scaling / model.fx();//内参矩阵中的　f/dx
  float constant_y = unit_scaling / model.fy();//内参矩阵中的　f/dy
  float bad_point = std::numeric_limits<float>::quiet_NaN();

  sensor_msgs::PointCloud2Iterator<float> iter_x(*cloud_msg, "x");
  sensor_msgs::PointCloud2Iterator<float> iter_y(*cloud_msg, "y");
  sensor_msgs::PointCloud2Iterator<float> iter_z(*cloud_msg, "z");
  const T* depth_row = reinterpret_cast<const T*>(&depth_msg->data[0]);
  int row_step = depth_msg->step / sizeof(T);
  for (int v = 0; v < (int)cloud_msg->height; ++v, depth_row += row_step)
  {
    for (int u = 0; u < (int)cloud_msg->width; ++u, ++iter_x, ++iter_y, ++iter_z)
    {
      T depth = depth_row[u];

      // Missing points denoted by NaNs
      if (!DepthTraits<T>::valid(depth))
      {
        if (range_max != 0.0)
        {
          depth = DepthTraits<T>::fromMeters(range_max);
        }
        else
        {
          *iter_x = *iter_y = *iter_z = bad_point;
          continue;
        }
      }

      // Fill in XYZ
      *iter_x = (u - center_x) * depth * constant_x;//这句话计算的原理是什么，通过内外参数矩阵可以计算
      *iter_y = (v - center_y) * depth * constant_y;//这句话计算的原理是什么，通过内外参数矩阵可以计算
      *iter_z = DepthTraits<T>::toMeters(depth);
    }
  }
}

} // namespace depth_image_proc

#endif