To generate point cloud data we need depth image and the intrinsic matrix of the depth camera. This package depends on depth_image_proc (source: https://github.com/ros-perception/image_pipeline/tree/melodic/depth_image_proc) to finish this part.
In the launchfile we first launch a depth_image_proc node, making it submit to the topic /realsense/depth/camera_infowhich contains the intrinsic matrix of th camera and the topic /realsense/depth/image which contains the raw depth image.
The vision node subscribes to the published point cloud and handles it with pcl::VoxelGrid filter. Here we set the leaf size of 0.2 meter to downsample the data.
Because the coordinate system of the camera is different from the one used in the mapping part, the point cloud data is rotated by 90 degree around the y-axis and -90 degree around the z-axis.