OrinNX上部署前端GPU加速的ORB_SLAM2_CUDA

诗筱涵

已于 2024-08-30 00:11:04 修改

阅读量315

点赞数 18

分类专栏：无人机-SLAM 无人机-SLAM-D435i 文章标签： ORBSLAM gpu

于 2024-08-29 22:38:16 首次发布

本文链接：https://blog.csdn.net/sinat_16643223/article/details/141690162

版权

无人机-SLAM 同时被 2 个专栏收录

191 篇文章 45 订阅

订阅专栏

无人机-SLAM-D435i

91 篇文章 30 订阅

订阅专栏

ORB_SLAM2_CUDA( https://github.com/thien94/ORB_SLAM2_CUDA )的编译运行环境和普通ORB_SLAM2/ORB_SLAM3的区别在于opencv和cv_bridge上，需要做些改动，其他的依赖项都是一样的。
基础opencv环境其实和vins-fusion-gpu是一样的，都是需要编译安装带有CUDA模块的opencv。
我这里是在部署了vins-fusion-gpu的OrinNX上进行ORB_SLAM2_CUDA的部署，opencv装的是3.4.1，编译时带上了CUDA相关选项。

opencv部署

新建一个opencv341文件夹，在里面下载3.4.1的opencv和opencv_contrib,使得opencv和opencv_contrib同在opencv341文件下。

git clone -b 3.4.1 https://github.com/opencv/opencv.git
git clone -b 3.4.1 https://github.com/opencv/opencv_contrib

然后在opencv文件夹下新建一个build文件夹，并运行下面命令

cmake -D CMAKE_BUILD_TYPE=RELEASE \
        -D CMAKE_INSTALL_PREFIX=/usr/local \
        -D WITH_CUDA=ON \
        -D CUDA_ARCH_BIN=7.2 \
        -D CUDA_ARCH_PTX="" \
        -D ENABLE_FAST_MATH=ON \
        -D CUDA_FAST_MATH=ON \
        -D WITH_CUBLAS=ON \
        -D WITH_LIBV4L=ON \
        -D WITH_GSTREAMER=ON \
        -D WITH_GSTREAMER_0_10=OFF \
        -D WITH_QT=ON \
        -D WITH_OPENGL=ON \
        -D CUDA_NVCC_FLAGS="--expt-relaxed-constexpr" \
        -D WITH_TBB=ON \
        -D OPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules \
         ../

OPENCV_EXTRA_MODULES_PATH参数后面的路径要和实际opencv_contrib的路径对应在NX上 CUDA_ARCH_BIN=7.2 ，在其他平台上这个参数值需要改 CMAKE_INSTALL_PREFIX对应sudo make install的时候实际的安装路径，修改后面的路径可以opencv所安装的文件夹。

上面命令运行完后接着运行下面命令

make -j8

sudo make install

安装好后，打开jtop，可以看到opencv 3.4.1的with CUDA后面显示的yes，代表我们装好了带有CUDA模块的opencv3.4.1，可以进行GPU加速。

输入图片说明

cv_bridge

下源码编译安装cv_bridge

    mkdir -p cv_bridge_ws/src
    cd cv_bridge_ws/src
    git clone https://gitee.com/bingobinlw/cv_bridge    
    cd ..

将opencv编译生成的build文件路径加入到CMakeLists.txt，个人感觉这步不加也可以应该，因为系统可以自动找到opemcv3.4.1了，保险起见还是加上。

set(OpenCV_DIR "your-path/opencv/build")

如果有find_package(OpenCV 4 QUIET)需要改为find_package(OpenCV 3 QUIET)，总之需要保证cv_bridge是寻找链接的opencv3.4.1，而不是opencv4

输入图片说明

catkin_make

把cv_bridge的功能包路径添加到bashrc里面，使得系统可以自动找到cv_bridge功能包

    echo "source ~/cv_bridge_ws/devel/setup.bash" >> ~/.bashrc
    source ~/.bashrc

Pangolin部署

Pangolin起可视化作用

git clone -b v0.6 https://gitee.com/maxibooksiyi/Pangolin

然后装些Pangolin的依赖，保险起见我没有直接运行脚本装，怕他给我更新了什么软件，其实下面这些装的时候大部分已经装了

sudo apt install libgl1-mesa-dev
sudo apt install libglew-dev
sudo apt install libpython2.7-dev
sudo apt install pkg-config
sudo apt install libegl1-mesa-dev libwayland-dev libxkbcommon-dev wayland-protocols

然后编译安装

cd Pangolin
mkdir build
cd build
cmake .. #有的地方写 cmake -DCPP11_NO_BOOSR=1 .. 也可以，#待验证区别
make
sudo make install
sudo ldconfig #这步容易漏掉

eigen安装

可以二进制安装

sudo apt-get install libeigen3-dev

编译ORB_SLAM2_CUDA

git clone https://github.com/thien94/ORB_SLAM2_CUDA

运行build.sh
会遇到下图所示报错

参考https://blog.csdn.net/For_Air_/article/details/135707182

把LoopClosing.h里面的

typedef map<KeyFrame*,g2o::Sim3,std::less<KeyFrame*>,
        Eigen::aligned_allocator<std::pair<const KeyFrame*, g2o::Sim3> > > KeyFrameAndPose;

改为：

    typedef map<KeyFrame*,g2o::Sim3,std::less<KeyFrame*>,
           Eigen::aligned_allocator<std::pair<KeyFrame *const, g2o::Sim3> > > KeyFrameAndPose;

输入图片说明

接着编译ros部分

需要先在.bashrc里面添加这句或者在运行build_ros.sh的终端运行这一句，不然编译会说找不到ORB_SLAM2_CUDA这个功能包

export ROS_PACKAGE_PATH=${ROS_PACKAGE_PATH}:/home/nvidia/ORB_SLAM2_CUDA/Examples/ROS

接着运行build_ros.sh

输入图片说明

参考 https://blog.csdn.net/weixin_44457432/article/details/130380189
cmakelists里加一句ADD_COMPILE_OPTIONS(-std=c++14 )

输入图片说明

再运行build_ros.sh编译完成

输入图片说明

ORB_SLAM2_CUDA运行

在ORB_SLAM2_CUDA/Examples/ROS/ORB_SLAM2_CUDA/conf下添加一个RealSense_D435i.yaml
这个RealSense_D435i.yaml的内容也是来自ORBSLAM3自己提供的D435i的配置文件

%YAML:1.0

#--------------------------------------------------------------------------------------------
# Camera Parameters. Adjust them!
#--------------------------------------------------------------------------------------------
File.version: "1.0"

Camera.type: "PinHole"

# Rectified Camera calibration (OpenCV)
Camera.fx: 382.613
Camera.fy: 382.613
Camera.cx: 320.183
Camera.cy: 236.455

# Kannala-Brandt distortion parameters
Camera.k1: 0.0034823894022493434
Camera.k2: 0.0007150348452162257
Camera.p1: -0.0020532361418706202
Camera.p2: 0.00020293673591811182

# Right Camera calibration and distortion parameters (OpenCV)
Camera2.fx: 382.613
Camera2.fy: 382.613
Camera2.cx: 320.183
Camera2.cy: 236.455

# Kannala-Brandt distortion parameters
Camera2.k1: 0.0034003170790442797
Camera2.k2: 0.001766278153469831
Camera2.p1: -0.00266312569781606
Camera2.p2: 0.0003299517423931039

Stereo.b: 0.0499585

# stereo baseline times fx ,Camera.fx * Stereo.b
Camera.bf: 19.13

# Camera resolution
Camera.width: 640
Camera.height: 480

# Camera frames per second 
Camera.fps: 30

# Color order of the images (0: BGR, 1: RGB. It is ignored if images are grayscale)
Camera.RGB: 1

# Close/Far threshold. Baseline times.
Stereo.ThDepth: 40.0

# Transformation from body-frame (imu) to left camera
Tbc: !!opencv-matrix
   rows: 4
   cols: 4
   dt: f
   data: [1.0, 0.0, 0.0, -0.005,
          0.0, 1.0, 0.0, -0.005,
          0.0, 0.0, 1.0, 0.0117,
          0.0, 0.0, 0.0, 1.0]

# Do not insert KFs when recently lost
IMU.InsertKFsWhenLost: 0

# IMU noise (Use those from VINS-mono)
IMU.NoiseGyro: 1e-3 # 2.44e-4 #1e-3 # rad/s^0.5
IMU.NoiseAcc: 1e-2 # 1.47e-3 #1e-2 # m/s^1.5
IMU.GyroWalk: 1e-6 # rad/s^1.5
IMU.AccWalk: 1e-4 # m/s^2.5
IMU.Frequency: 200

#--------------------------------------------------------------------------------------------
# ORB Parameters
#--------------------------------------------------------------------------------------------
# ORB Extractor: Number of features per image
ORBextractor.nFeatures: 1250

# ORB Extractor: Scale factor between levels in the scale pyramid 	
ORBextractor.scaleFactor: 1.2

# ORB Extractor: Number of levels in the scale pyramid	
ORBextractor.nLevels: 8

# ORB Extractor: Fast threshold
# Image is divided in a grid. At each cell FAST are extracted imposing a minimum response.
# Firstly we impose iniThFAST. If no corners are detected we impose a lower value minThFAST
# You can lower these values if your images have low contrast			
ORBextractor.iniThFAST: 20
ORBextractor.minThFAST: 7

#--------------------------------------------------------------------------------------------
# Viewer Parameters
#--------------------------------------------------------------------------------------------
Viewer.KeyFrameSize: 0.05
Viewer.KeyFrameLineWidth: 1.0
Viewer.GraphLineWidth: 0.9
Viewer.PointSize: 2.0
Viewer.CameraSize: 0.08
Viewer.CameraLineWidth: 3.0
Viewer.ViewpointX: 0.0
Viewer.ViewpointY: -0.7
Viewer.ViewpointZ: -3.5
Viewer.ViewpointF: 500.0

在ORB_SLAM2_CUDA/Examples/ROS/ORB_SLAM2_CUDA/launch下添加一个ros_mono_d435i.launch

<?xml version="1.0"?>
<launch>

    <arg name="vocabularty_path" default="$(find ORB_SLAM2_CUDA)/../../../Vocabulary/ORBvoc.txt" />
    <arg name="camera_setting_path" default="$(find ORB_SLAM2_CUDA)/conf/RealSense_D435i.yaml" />
    <arg name="bUseViewer" default="true" />
    <arg name="bEnablePublishROSTopic" default="true" />

    <node name="ORB_SLAM2_CUDA" pkg="ORB_SLAM2_CUDA" type="Mono" output="screen" 
    	args="$(arg vocabularty_path) $(arg camera_setting_path) $(arg bUseViewer) $(arg bEnablePublishROSTopic)">
  	</node>

</launch>

把ORB_SLAM2_CUDA/Examples/ROS/ORB_SLAM2_CUDA/src/ros_mono.cc里订阅的图像话题名称改为D435i的图像话题名称，改完保存后需要重新运行build_ros.sh进行编译

    ros::Subscriber sub = nodeHandler.subscribe("/camera/infra1/image_rect_raw", 1, &ImageGrabber::GrabImage,&igb);

再运行下面命令可以将ORB_SLAM2_CUDA起起来

roslaunch /path/to/ORB_SLAM2_CUDA/Examples/ROS/ORB_SLAM2_CUDA/launch/ros_mono_d435i.launch

注意需要先在.bashrc里面添加这句或者在运行上述命令的终端里运行这一句，不然会找不到ORB_SLAM2_CUDA这个功能包

export ROS_PACKAGE_PATH=${ROS_PACKAGE_PATH}:/home/nvidia/ORB_SLAM2_CUDA/Examples/ROS

接着再运行D435i或者播放D435i的bag包，可以将基于D435i的ORB_SLAM2_CUDA运行起来
从jtop也可以看到gpu有明显被占用

输入图片说明

终端也实时打印每帧的耗时情况

输入图片说明

诗筱涵

关注

18
点赞
踩
11

收藏

觉得还不错? 一键收藏
0
评论
OrinNX上部署前端GPU加速的ORB_SLAM2_CUDA

ORB_SLAM2_CUDA的编译运行环境和普通ORB_SLAM2/ORB_SLAM3的区别在于opencv和cv_bridge上，需要做些改动，其他的依赖项都是一样的。基础opencv环境其实和vins-fusion-gpu是一样的，都是需要编译安装带有CUDA模块的opencv。我这里是在部署了vins-fusion-gpu的OrinNX上进行ORB_SLAM2_CUDA的部署，opencv装的是3.4.1，编译时带上了CUDA相关选项。
复制链接

扫一扫

专栏目录