semantic_slam环境配置

orb_slam2 + 语义分割 + 八叉树地图,
github地址：https://github.com/floatlazer/semantic_slam

根据github上环境的版本要求，选择在ubuntu16.04的docker容器中搭建

首先pull了一个ubuntu16.04 + python2.7 + opencv3.3.0的镜像，在这个基础上搭建环境

1.download semantic slam的code
在$HOME下

$ mkdir catkin_ws
$ cd catkin_ws
$ mkdir src
$ cd src
$ git clone https://github.com/floatlazer/semantic_slam.git

2.安装orb slam2的环境
参考了https://www.jianshu.com/p/dbf39b9e4617
这中间涉及到ros, pangolin, eigen等的安装

进入sematic_slam文件夹中的ORB_SLAM2文件夹，build成功即可

3.安装语义分割相关
按照semantic slam的github中的要求

$ sudo apt-get install ros-kinetic-openni2-launch

4.安装八叉树地图相关

sudo apt-get install ros-kinetic-octomap-ros
sudo apt-get install ros-kinetic-octomap-msgs
sudo apt-get install ros-kinetic-octomap-server
sudo apt-get install ros-kinetic-octomap-rviz-plugins

5.安装语义分割相关

pip install torch-0.4.0-cp27-cp27mu-linux_x86_64.whl #1.0之前的版本要在官网上下载
pip install torchvision==0.2.0
pip install numpy
pip install future
pip install matplotlib

然后github上说依赖pytorch-semseg
打开pytorch-semseg中的requirements.txt, 看到里面还有一些依赖要安装

pip install scikit-image
pip install opencv-python==3.3.0.10
pip install tqdm==4.11.2
pip install cython
pip install pydensecrf
pip install tensorboardX
pip install scipy==0.19.0
pip install imgaug

然后安装semantic_slam包

$ rosdep install semantic_slam
ERROR: the following packages/stacks could not have their rosdep keys resolved
to system dependencies:
semantic_slam: Cannot locate rosdep definition for [catkin]

这一步会报找不到相关依赖的错误
命令改为：

$ rosdep install --from-paths semantic_slam --ignore-src --rosdistro=kinetic -r -y
#All required rosdeps installed successfully

这些都安装好之后，到~/catkin_ws下catkin_make

开始启动launch:

roslaunch semantic_slam semantic_mapping.launch

会报如下错误：
ImportError: No module named google.protobuf.internal
解决方法：

pip install protobuf

再启动launch:
会报错：AttributeError: ‘module’ object has no attribute ‘_cuda_getDevice’
Traceback (most recent call last):
File “/home/xx/catkin_ws/src/semantic_slam/semantic_cloud/src/semantic_cloud.py”, line 295, in
main(sys.argv)
File “/home/xx/catkin_ws/src/semantic_slam/semantic_cloud/src/semantic_cloud.py”, line 288, in main
seg_cnn = SemanticCloud(gen_pcl = True)
File “/home/xx/catkin_ws/src/semantic_slam/semantic_cloud/src/semantic_cloud.py”, line 123, in init
state = torch.load(model_path)

这是因为它的weight, 也就是pth文件，是用GPU训练出来的，
而docker里面没有GPU, 要把这个pth转为cpu模式：
把这个 state = torch.load(model_path)
改为：

state = torch.load(model_path, map_location={'cuda:0': 'cpu'})

到~/catkin_ws下

$ rm -rf build devel
$ catkin_make

再试一次

$ roslaunch semantic_slam semantic_mapping.launch
[ INFO] [1655366319.121127159]: Semantic octomap generator [max fusion]
Generate semantic point cloud [max fusion].
Setting up CNN model...
Setting up ROS...
Ready.

另起一个窗口

$ rosbag play demo.bag

图建出来了