ROS2: GPU/NVEncoder Acceleration

[Owen-Liuyuxuan]

This PR is the ROS2 refinement of #11.

It is still under-test.

Basic idea

• Break ros2_v4l2_camera as two standalone packages. (If we further stress on efficiency instead of flexibility, we should consider merging them in the future).
• The ros2_v4l2_camera stays the same as in the main. But I do not recommend launching the camera drivers with image_transport (which automatically use CPU encoding).
• The new gpu_imgproc is the ROS2 refreshment of the ROS1 - GPU acceleration #11 .

Parameteres and Topics Changes:

Topics	Direction	Launch Parameters	Description
/image_raw	subscribe	${input_image_topic}	Input raw image with RGB8 encoding sensor_msgs/Image
/camera_info	subscribe	${camera_info_topic}	The camera calibration information sensor_msgs/CameraInfo
/image_raw/compressed	publish	${input_image_topic}/compressed	The compressed output sensor_msgs/CompressedImage
/image_rect	publish	${input_image_topic}/image_rect	The rectified raw image sensor_msgs/Image
/image_rect/compressed	publish	${input_image_topic}/image_rect/compressed	The output compressed rectified image sensor_msgs/CompressedImage

Parameters	default	Description
rect_impl	npp	select rectifier implementation from npp, opencv_cpu, opencv_gpu.
use_opencv_map_init	False	select implementation of rectifier mapping, from OpenCV or NPP
alpha	0.0	The free scaling parameter for rectification; see OpenCV Documentation
input_image_topic	image_raw	Input image topic, will also influence the output
camera_info_topic	camera_info	Input camera info topic

Test going to implement:

• As part of integration research, we will test the code on work on both x64 CPUs but also on Jetson machines. The base platform will be the Adlink ROSCube 580&58g.
• We are first going to make sure we are actually using GPU/NVEncoder to compress the data.
• Then we will benchmark the results.
• Hopefully, all ROSCube/Orin-related projects should be using the GPU encoder instead of the CPU encoder.

Quick Testing Use of this PR:

# Beforehand, set up ROS2 download/paths and CUDA download/paths;
# Then under a workspace
cd  src
git clone --branch galactic https://github.com/tier4/ros2_v4l2_camera.git
git clone jetson_process
cd ../
rosdep install --from-paths src --ignore-src -r -y
colcon build --symlink-install --cmake-args -DCMAKE_BUILD_TYPE=Release

Testing:

• Firstly, it works on ROSCube 58g and it boosts up NVJPEG.
• Comparing the decompressing capability of image_transport and our PR. Testing against 20Hz of 1920*1440 image streams. ROS2 image_transport will use 60%-80% of CPU. This PR uses 20%-30% of CPU. We can see a significant reduction on the CPU time.

[Owen-Liuyuxuan]

With the new commit, I have tested with x86-64 computers and the decompress process works well. (At least it is not dying and libturbojpeg works fine)

[Owen-Liuyuxuan]

https://github.com/tier4/accelerated_image_processor/tree/perf/modify_cuda_utilization によってcloseします。